CellCog Release Notes

Video generation on CellCog now costs 49% less per second. Since October 15, 2026, every Seedance 2.5 render at 480p, 720p and 1080p costs about half what it did, from the very first shot. It is the same model with the same output: we moved it to a cheaper host, and because CellCog credits follow what the work costs us, the whole saving passes to you.

The second change matters more for anyone making real films: drafts. Your AI employee can now render a 480p draft, let you judge the take, and then finish that exact take at 1080p. Together, the two changes cut the cost of a 1080p film by about three quarters.

Half price, from the first shot

If your habit was 720p, you now pay about half. If you rendered 1080p finals directly, those cost about half too. Nothing about the clip changes: Seedance 2.5 still generates 4 to 30 seconds at 24 fps with audio included, from up to 50 reference images, videos and audio clips. A backup host takes over automatically if the main one is briefly down, so a job does not fail because one provider had a bad minute.

Drafts: judge the take, then finish that same shot

Video generation is a roll of the dice. Before this change, the only way to a 1080p final was to render at 1080p, and every re-render rolled a new take. A scene that needed three tries meant paying for three full 1080p renders.

A draft breaks that. It renders at 480p and costs what any 480p render costs, a little under a fifth of the 1080p finish. You look at it. If the take is wrong, you draft again cheaply. When a take is right, the finish renders the identical shot at 1080p: same composition, same motion, same audio.

A draft locks the prompt, references, duration, aspect ratio, seed and audio, and remains valid for 7 days in the chat that made it. The finish renders the same take at 1080p, 24 fps. If the very first take is the keeper, draft plus finish costs about 18% more than going straight to 1080p at today’s price. Drafts pay off from the second take on.

What a 1080p final costs now

A 1080p final made with drafts is 39% cheaper when the first take is the keeper, 65% cheaper on the second take, 74% cheaper on the third, and 80% cheaper on the fifth. A two-minute 1080p film built from eight scenes with three takes each costs about 74% less than the same film did before.

Your AI employee decides when to draft

You do not have to learn a new step. In a blind test, a plain request for a short coffee-shop clip and something to approve before the 1080p final led the agent to make a draft on its own, quote the finish cost, and deliver the same take at 1080x1920 after approval. Drafts are for 1080p finals; for a 720p deliverable, CellCog renders 720p directly at half the old price. A draft is made at the final length, and it expires after 7 days in the chat that made it.

Also back: fictional characters as references

Character references of fictional people work again with photoreal, painted or illustrated faces, so a character stays the same from the first scene of a long film to the last. For real people, your own likeness works; anyone else needs their consent; public figures and minors are never allowed.

Caveats

The percentages compare CellCog’s price today with its price before October 15, 2026, for 16:9 clips without a reference video. Jobs with reference video bill differently, and the saving can differ. If the main host is down, a job runs on the backup at the old price, and the backup refuses face references; the agent tells you and retries. Plans and credit rates are available on cellcog.ai/pricing.

CellCog agents can now hand work to subagents. A regular chat or an AI employee writes a brief, starts a helper agent (another CellCog chat, or Claude Code, Codex or Cursor running on your own computer under your own login), and then rests instead of waiting. When the helper finishes, it calls a wake link, CellCog wakes the parent, and the parent reads the work and marks it reviewed. You can see that last step: every finished subagent carries a small dot until your agent has actually looked at its output.

Subagents shipped in three steps between October 10 and October 12, 2026, and all of it is live. This post covers how a hand-off works, what it costs, when your agent uses a subagent and when it does not, and how other agent products handle the same hand-off, quoted from their own docs.

KEY POINTS

A CellCog chat or AI employee can now hand a piece of work to a subagent, rest while it works, and pick up again the moment it finishes. Shipped October 10 to 12, 2026, all live.

Five kinds: a new CellCog chat, or Claude Code, Codex or Cursor running on your own computer under your own login, or anything that can call a web link.

Each subagent gets a wake link that works 24 hours, allows 3 calls and carries no data. A link that runs out unused wakes the parent once with a never-reported note.

A finished subagent shows a dot until the parent agent reads its output and marks it reviewed. Opening it yourself never clears the dot.

Coding subagents use your own Claude, ChatGPT or Cursor plan first; a CellCog subagent uses credits like any chat. You pay for the work, not the hire.

Your agent launches a subagent with no more access than it has itself. Starting a coding subagent follows your approval level, and anything that needs your approval comes back to you.

§ 01 What shipped

Subagents shipped between October 10 and October 12, 2026:

• October 10 — Subagents: a chat or an AI employee starts a child agent and rests until it reports back. Five kinds: CellCog, Claude Code, Codex, Cursor and Other.
• October 11 — The strip above the message box, Check now for your coding tools, usage readings, and the review step.
• October 12 — One state per chip (running, finished, reviewed) and a written test every agent follows for when to use a subagent and when not to.

§ 02 Five kinds of subagent

• CellCog: A new CellCog chat, started by the parent. Nothing; CellCog sets it up.
• Claude Code: On your own computer, through Cowork. CellCog Desktop connected; Claude Code installed and signed in.
• Codex: On your own computer, through Cowork. CellCog Desktop connected; Codex installed and signed in.
• Cursor: On your own computer, through Cowork. CellCog Desktop connected; the Cursor command line agent installed and signed in.
• Other: Anything that can call a web link. Whatever that tool needs.

Claude Code, Codex and Cursor run under your own login with those products and use your plan with them. CellCog does not run them in the cloud. Starting one is a command on your computer, so it goes through your approval level like any other Cowork command.

§ 03 How a hand-off works

1. The parent creates the subagent. CellCog gives it an id such as SA-7KQ2MX and a wake link.
2. The parent writes the wake link into the brief, with a plain instruction: call this link when you are done.
3. The parent starts the subagent and records where its work will land: a log file, a branch, a chat.
4. The parent rests. It is not running, and not spending, while the subagent works.
5. The subagent finishes and calls the wake link. CellCog wakes the parent with a short note naming it.
6. The parent reads the result itself, from the log, the changes or the CellCog chat, marks it reviewed and carries on.

The wake link works for 24 hours and can be called at most 3 times; every call wakes the parent. It carries no data. It only says “I was called”, and the parent reads the actual work from where the subagent left it, so nothing the helper writes into a link can steer the parent. Long work gets a fresh link for each round.

A subagent can crash, run out of room or hit a plan limit before it reports. If its link runs out unused, CellCog wakes the parent once with a note that the subagent never reported, so the work is never silently lost. The parent can also set an expected time; past it, the note arrives early, and a late call still wakes the parent.

§ 04 Reviewed, not just done

The part we care most about is the dot. When a subagent finishes, or never reports, a small dot appears on it in the chat and on the Subagents page. It means “not reviewed by your agent yet.” Only the agent that started it can clear it, by reading the output and marking it reviewed. Opening the subagent, the chat or the Subagents page yourself never clears it, so the dot stays an honest signal that your agent looked at the work before it told you it was done.

In the chat, a thin strip above the message box shows each subagent by name. A logo pulses while it works, a dot appears when it finishes, and a green tick replaces the dot once it is reviewed. The strip goes away when every subagent on it has its tick. If the same subagent reports again, the dot comes back until that round is reviewed too.

Here is an illustrative case. An AI software engineer is asked to upgrade a payments library across a large codebase. First a CellCog subagent reads the library’s changelog and lists the breaking changes. Then Claude Code and Codex each migrate a separate set of modules, each in its own working copy so no two edit the same files, while the engineer rests. Last, a fresh-eyes review subagent reads the combined diff. As each one reports, the engineer reads its output, marks it reviewed, and only then tells you the upgrade is ready.

§ 05 When your agent uses one, and when it does not

Subagents are not a feature where more agents means better work. Every CellCog agent now carries the same written test.

USE A SUBAGENT

• The work splits into pieces that each return a small, checkable result.
• A long build, test run or wait can run while the parent rests.
• Noisy reading (logs, a codebase, many sources) has a one-page answer.
• A fresh-eyes review of finished work.

DO IT YOURSELF

• The work is sequential or tightly coupled.
• It is the decision itself.
• The plan is still moving.
• The brief would take as long as the change, or two writers would touch the same file.

The parent decides and integrates; the child executes inside a fence. Decisions, design and anything that needs your approval stay with the parent.

§ 06 Your plans and your keys

Open Subagents under Apps in the sidebar. A strip at the top shows whether each kind is ready: your computer, CellCog, Claude Code, Codex and Cursor. Click a coding tool for its four setup steps. Check now runs a short, read-only check on your computer for all three tools at once: whether each is installed, whether you are signed in, and whether an API key is set in your terminal. The key never leaves your computer.

A subagent uses your plan first, your Claude, ChatGPT or Cursor subscription. A key set in your shell profile can quietly take over from a plan and bill your API account instead, so for a subagent’s process that key is set aside. If you have both, a switch, Use my API key when my plan’s limit is reached, decides what happens at the limit. It is off at first: the subagent stops and your agent tells you, waits for the reset or uses another kind.

Before choosing a tool, your agent reads each one’s newest usage reading, how much of the plan’s limits was in use as the tool itself reported it, so your agent can pick a tool with room left. Claude Code reports it in its own run output and Codex through Check now; Cursor publishes no usage data.

For CellCog subagents, the first one creates an API key named Subagents for you in that organization. It shows on your API keys page and under Connectors, can be rotated but not deleted, and does not count toward your active-key limit. Your agent uses it from its own workspace and never sees or prints it.

§ 07 What it costs

You pay for the work, not the hire; the cost depends purely on how much work you assign. A CellCog subagent is a CellCog chat and uses credits like any chat. Claude Code, Codex and Cursor use your own plan with those products, not CellCog credits. Each time a subagent wakes the parent, the parent does a normal run, paid by the account that owns the chat or the AI employee. Current plans are on the pricing page.

§ 08 How other agent products hand off work

Several agent products now let a parent agent hand work to helpers. Here is how each describes the moment a helper finishes, in its own docs.

• CellCog — CellCog chats, or Claude Code, Codex and Cursor under your own login; CellCog, or your own computer; it calls a wake link; the parent reads the work and marks it reviewed, and you see a dot until it does.
• Claude Code — Its own subagents, each with its own context window; inside one Claude Code session; the subagent returns a summary to the main conversation.
• Codex — Parallel subagents; the Codex app and CLI; the main thread collects the results; you can open each agent thread.
• Devin CLI — Its own subagents, in the foreground or background; inside the Devin CLI; the parent reads the result and summarizes it for you.
• Manus Wide Research — Many parallel agents, one item each; its page does not say; the main agent collects and assembles the results.
• Grok Bot — Cursor Cloud Agents, for coding work; separate computers under Cursor’s Cloud Agent controls; the Bot attaches the files cited in the Cloud Agent’s final report to its reply.

Each one does something well. Anthropic’s page explains the core value plainly: a subagent takes a noisy side task, “does that work in its own context and returns only the summary.” It also marks every returning report so that “instructions or approval claims inside the report are the subagent’s words and carry no authority from you.” OpenAI is candid that “Subagent workflows consume more tokens than comparable single-agent runs.” Devin comes closest to our review step: “When a subagent finishes, the parent agent reads the result and summarizes the key findings and actions for you.” xAI’s Grok Bot hands coding work to Cursor: “Grok Bot can delegate coding tasks to separate computers under your existing Cloud Agent controls.”

Where CellCog differs is in three places. Its helpers include the coding tools you already pay for, running on your own computer under your own login. The parent rests until a wake link calls it back instead of holding a session open. And whether the parent has reviewed each helper’s work is something you can see, not something you take on trust. None of the five pages we read describes a review mark like the dot.

§ 09 Built for an organization of AI employees

An AI employee’s subagents belong to the employee, not to one conversation. If the employee has started a fresh working session since launching them, the note still reaches it. The Subagents page groups every chat and AI employee that started subagents in your current organization, newest first, with each subagent’s status, what the parent recorded about it, and when it last reported. A waiting subagent has a Cancel button, and cancelling stops its wake links at once.

Your agent launches a subagent with no more access than it has itself. Starting a coding subagent is a command on your computer and follows your approval level; what it may do after that is set when your agent launches it, and anything beyond that comes back to your agent, and to you when it needs approval.

§ 10 The honest caveats

• Claude Code, Codex and Cursor subagents need the CellCog Desktop app connected and the tool installed and signed in. They run only while your computer is awake; one that stops before reporting sends the never-reported note.
• Codex’s usage reading has not yet been proven against a live install, and Cursor publishes no usage data at all.
• The wake link carries no data and allows 3 calls, so multi-round work needs a fresh link per round.
• The safety model is your approval level plus the doctrine above, not a lock in the code. On machines that hold credentials, our agents are told to launch coding tools with a clean environment and a sandbox, and to ask the parent for anything beyond that box.
• Subagents are visible to the agent that started them and on the Subagents page of the current organization.

A CellCog AI software engineer can hand pieces of a job to Claude Code, Codex or Cursor on your own plans, rest while they work, and read every result before it tells you the job is done. Try it free, no credit card needed.

FREQUENTLY ASKED

Q1 Do I have to do anything when a subagent finishes?

No. CellCog wakes the parent and it continues on its own. If a chat parent needs your input, it asks you as usual.

Q2 Does a subagent use CellCog credits?

A CellCog subagent is a CellCog chat and uses credits like any chat. Claude Code, Codex and Cursor use your own plan with those products. Each wake is a normal run for the parent. Plans are on cellcog.ai/pricing.

Q3 Can an AI employee use subagents?

Yes. An AI employee’s subagents belong to the employee, so a note reaches it even if it has started a fresh working session since launching them.

Q4 What if my computer is off?

Claude Code, Codex and Cursor subagents need your computer on and CellCog Desktop connected to start, and keep running only while it is awake. If one stops before reporting, the parent gets the never-reported note.

Q5 Can a subagent do more than the agent that started it?

Your agent launches a subagent with no more access than it has itself. Starting a coding subagent is a command on your computer and follows your approval level; what it may do after that is set at launch, and anything beyond that comes back to your agent, and to you when it needs approval.

KEY POINTS

5 · 6 min full read

• CellCog runs a remote MCP server at https://cellcog.ai/mcp, live since September 17, 2026. Claude, Claude Code, Codex, Cursor and other clients that speak MCP can connect to it.
• You sign in once with your own CellCog account on a consent page and click Approve. There is no API key and no shared team token; each person connects as themselves.
• Eight tools let the assistant hand CellCog a brief and get finished work back: research, code, dashboards and apps, PDFs and decks, video, images, audio and spreadsheets. Each brief runs as an agent job, not a single model call.
• The assistant checks back about once a minute and gets progress notes while the job runs, then the files once, each with a Download link valid for 30 days.
• It is billed like a web chat, from your credits or your organization’s shared pool. Not yet: hiring or managing AI employees, or acting on your computer or browser, through the connector.

§ 01 The reverse of what shipped in August

On August 31 we shipped the other direction: connect your own MCP servers to CellCog so your agents can call your tools. This update turns CellCog itself into a server. Your assistant stays where you work, and CellCog becomes the place it sends the jobs that need an agent: research, code, dashboards and apps, PDFs and decks, video, images, audio and spreadsheets.

What comes back is the result of an agent job, not a single model call. A brief for a launch video comes back as the script, the clips and a check of the result together.

§ 02 Connect in one step

How to add CellCog to your assistant (October 7, 2026):

• Claude web or Desktop: Settings, Connectors, Add custom connector, paste the address, Approve, then start a new chat
• Claude Code: claude mcp add --transport http cellcog https://cellcog.ai/mcp
• Codex: codex mcp add cellcog --url https://cellcog.ai/mcp, then codex mcp login cellcog
• Cursor: The one-click install button in CellCog’s MCP / API app, or add the address as a remote server
• Any other MCP client: Add https://cellcog.ai/mcp as a remote (streamable HTTP) server

The first use opens CellCog’s consent page, which names your assistant (“Claude wants to use your CellCog account”) and lists three permissions in plain words. The assistant can create chats and send messages in your account, read chat status and results, and open files the agent produced, with links valid for 30 days. You click Approve once. There is no API key to copy and no shared team token; every person signs in as themselves.

Inside CellCog, the MCP / API item in the left sidebar holds the address, the setup for each assistant and the apps you have connected. New accounts also see CellCog MCP as a choice on their first screen. The CellCog MCP server page sets the connector beside the other way in: an API key for scripts and agents you build yourself.

§ 03 What an assistant can do with it

Eight tools sit behind the address:

• Create chat: starts a CellCog chat from a brief, with the agent and depth you choose.
• Send message: a follow-up in the same chat, in context.
• Upload files: an upload window for files on your computer, up to 10 files of 25 MB each, kept 24 hours.
• Get status and Get result: check on a job and collect what it made.
• Stop chat: stops a running job.
• List chats and Account: recent chats and the account the assistant is signed in as.

§ 04 How a job comes back

Agent jobs take time, so the connector reports progress instead of making the assistant wait. While a job runs, Get result returns the agent’s progress notes since the last check and suggests calling again in about a minute; video jobs take 10 to 40 minutes. When the job finishes, Get result returns the final message, the files and the credits used once, with a one-line summary such as “Finished: 3 files”.

Every file carries a Download link valid for 30 days. Clients that render inline views show players, images, PDF page previews and spreadsheet formula tables in the conversation; others show the links.

§ 05 Three weeks of testing

Timeline of the CellCog MCP server: live on September 17, progress notes and one-shot results on September 19, eight tools and file uploads on September 21, the MCP / API app on October 2, and this update on October 7.

§ 06 How other MCP servers compare

Several products now run their own MCP server. What a call does is where they differ.

Vendor-hosted MCP servers, per each company’s own docs (read October 7, 2026):

• CellCog: cellcog.ai/mcp, OAuth with your CellCog account, starts an agent job from a brief, billed same as a web chat
• Higgsfield: mcp.higgsfield.ai/mcp, OAuth no API key, generates images, video, characters and audio, MCP always uses credits
• Relevance AI: mcp.relevanceai.com, OAuth, triggers your Relevance agents; long runs start async and are polled
• Zapier: mcp.zapier.com/api/v1/connect, OAuth or a connection token, runs the app actions you configured

Higgsfield’s page is direct about credits: “everything generated through MCP deducts credits, regardless of your plan.” Relevance AI describes the same pattern we use for long work: “The trigger returns immediately with a conversation ID, which you then poll to check status and retrieve results.” Zapier’s strength is reach into your apps: “Zapier holds the app credentials, so your MCP client never handles a third-party API key.”

Brief CellCog from the assistant you already use.

Connect once at cellcog.ai/mcp, sign in with your CellCog account, and hand off the jobs that need an agent. You pay for the work, not the hire, and the cost depends on how much work you assign. Try it free, no credit card needed.

§ 07 The honest caveats

• A connected assistant cannot hire or manage AI employees, change your connectors, or act on your computer or browser.
• Some sandboxes, such as Claude Cowork on Team plans, cannot download the files themselves until an admin allows cellcog.ai and storage.googleapis.com. You can always open the links yourself.
• Clients that only support local (stdio) servers cannot connect; they use the CellCog plugin or the SDK.
• CellCog is a custom connector in Claude, added by address. It is not listed in Anthropic’s connector directory.

§ 08 Sources

• CellCog, MCP Connector Guide, read October 7, 2026.
• Higgsfield, What is Higgsfield MCP, read October 7, 2026.
• Relevance AI, MCP server, read October 7, 2026.
• Zapier, How Zapier MCP connections work, read October 7, 2026.

FREQUENTLY ASKED

Q1 How do I add CellCog to Claude?

In Claude web or Desktop, open Settings, then Connectors, choose Add custom connector, paste https://cellcog.ai/mcp, click Approve on CellCog’s consent page, then start a new chat. In Claude Code, run claude mcp add –transport http cellcog https://cellcog.ai/mcp.

Q2 How do I add CellCog to Codex or Cursor?

In Codex, run codex mcp add cellcog –url https://cellcog.ai/mcp, then codex mcp login cellcog. In Cursor, use the one-click install button in the MCP / API app in CellCog’s left sidebar, or add the address as a remote server.

Q3 What can an assistant do through the connector?

Create a chat from a brief, send follow-ups, upload files from your computer, check status, collect results, stop a running chat, list recent chats and read the account. Each brief runs as a CellCog agent job and returns files with Download links.

Q4 How long do the file links last?

30 days. Every file comes with a Download link, and clients that render inline views show players, images and page previews in the chat.

Q5 Is this the same as connecting my own MCP servers to CellCog?

No, it is the reverse. Since August 31 you can plug your own MCP servers into CellCog. This update is CellCog acting as an MCP server that other assistants connect to.

Your AI employees can now buy things for you. Spend lets an employee put an order together, ask you to approve the exact total, and pay from your own Link wallet once you say yes. You approve every purchase in Link, and the card it pays with works once, only for that purchase. The errands are ordinary: two copies of a book for the reading group, a train ticket, a conference pass, a part from a supplier. Until now an employee could research and prepare all of it and then had to stop at the checkout. Now it can finish the job.

KEY POINTS

Spend is live: a CellCog AI employee can buy things for you, paid from your own Link wallet, Stripe’s wallet for paying online. It is off for every employee until you turn it on.

You approve every purchase yourself, in Link, at the exact total and merchant your employee requested. CellCog sends no extra notification; the approval lives in Link.

On approval, Link issues a one-time card that works once, only at that merchant, for that amount. Your real card stays in Link and never reaches CellCog.

Each request waits 30 minutes for your answer. If the checkout total changes, your employee cancels and asks again with the new total.

Link connects once per organization and can serve every employee you own there. Each employee gets its own switch, optional caps and purchase history, all in the Spend app.

Not yet: hotels that hold a deposit at check-in, subscriptions, and currencies other than US dollars. Link’s agent wallet is for people in the US and Canada.

§ 01 How a purchase works

A purchase from ask to receipt:

1. Ask - You: In chat, by email or on a task: what to buy. Your employee buys only what you asked for or what is plainly part of work you gave it
2. Prepare - Your employee: In its own browser it fills the cart, picks shipping and reads the final total, tax and shipping included
3. Request - Your employee: It asks Link to approve that exact total at that merchant, with the items and a short note on why
4. Approve - You, in Link: Link notifies you in its own app. You have 30 minutes to approve or decline
5. Pay - Your employee: Link issues a one-time card for that amount and merchant; your employee completes the checkout and records the confirmation
6. See it - You: Every purchase and its status shows in the Spend app and as a line in the employee’s chat

If the total changes before payment, the approved card no longer fits, so your employee cancels and asks again with the new total. If you decline, it drops the purchase and does not ask again unless you tell it to.

§ 02 Built for an organization of employees

Spend was built for a team, not one assistant. You connect Link once per organization, and every employee you own there can use it, each with its own switch, its own optional caps and its own purchase history. Your research assistant can buy the report it needs while your operations manager has purchases switched off entirely.

What you control, per employee:

• Make purchases - Default Off: Nothing is bought until you turn it on for that employee
• Per purchase cap - Default No cap: A flight or a conference pass can cost more than any default we could pick
• Monthly cap - Default No cap: Add one if you want a ceiling; your approval in Link stays the real check
• Link connection - Default One per organization: A wallet connected in one organization never pays for another organization’s employees

Caps are optional on purpose. Your approval in Link decides every purchase; caps are an extra limit you add only if you want one.

§ 03 How other agents pay

Agents that buy things are arriving fast, and they do not all ask the same way. From each company’s own pages, read on October 4, 2026:

How four agents handle a purchase, per each company’s docs:

• CellCog AI employees - One-time card from your Link wallet - You approve every purchase in Link, at the exact total
• Meta Muse - Link one-time card (recommended), or a card saved on a merchant account - Muse “will always ask for your approval prior to completing a purchase”
• Grok Bot - A single-use card from Link - “You approve every spend request”, US users
• OpenAI dots - A card you saved on the merchant’s website - Required, and “may be given in advance when it specifically covers the purchase”

Link’s own documentation describes the same rail Spend uses: “Your agent creates a spend request, the customer approves it, and Link returns a one-time-use payment credential for that purchase” (Stripe docs). Our difference is the organization around it: one wallet, many employees, each switched on separately, all in one app.

Hand the errand to someone who can finish it.

An AI executive assistant can research, compare, prepare the order and, with Spend on, complete the purchase once you approve it in Link.

§ 04 The honest caveats

• Where it works. Spend fills the card where the card form sits on the merchant’s own site or a known hosted checkout. Some checkouts still need you for a bank check or a captcha; your employee fills everything else and sends you a link to finish that step.
• What it cannot buy yet. Hotels that charge at check-in or hold an incidentals deposit, subscriptions and renewals, and purchases in currencies other than US dollars.
• Where Link’s agent wallet runs. Per Stripe, “Agent payments are available to US and Canadian consumers.”
• Screenshots. Your employee can see the one-time card in its checkout screenshots. That card works once; your real card never reaches CellCog.
• Who is buying. You are. It is your money and your wallet. Order and refund questions go to the merchant; card questions go to Link or your bank.

§ 05 Getting started

1. Open Spend under More in the left sidebar and connect Link.
2. Open an employee’s Settings app and turn on purchases, with caps if you want them.
3. Ask for something: “Order two copies of Working in Public for the reading group.”

The full setup and every status are in the Spend guide.

§ 06 Sources

• CellCog, Spend guide.
• Stripe Docs, Enable agents to spend, read October 4, 2026.
• Meta Help Center, How Muse works with payments, read October 4, 2026.
• OpenAI Help Center, Dots privacy, security, and safety FAQs, read October 4, 2026.
• Grok Bot on X, August 28, 2026.

FREQUENTLY ASKED

Q1 How do I turn Spend on?

Open Spend under More in the left sidebar and connect Link once for your organization. Then open the employee’s Settings app and turn on purchases for that employee, with optional per-purchase and monthly caps.

Q2 Can my employee buy something I did not approve?

No. Every purchase needs your approval in Link, at the exact total and merchant, and the one-time card works once, only for that amount and merchant.

Q3 Can my employee see the card it pays with?

It can see the one-time card in checkout screenshots. That card works once, for this purchase only. Your real card never reaches CellCog.

Q4 What can it not buy yet?

Hotels that charge or hold a deposit at check-in, subscriptions and renewals, and purchases in currencies other than US dollars. For those, your employee prepares everything and you pay yourself.

Q5 What happens at a bank check or captcha?

Your employee fills in everything it can and sends you a link to its browser so you can finish that one step.

Q6 Who do I contact about an order or refund?

The merchant. It is your purchase from your wallet, and CellCog is not a party to it. Card questions go to Link or your bank.

Contacts is live: one shared record of every person your AI employees email, call or meet.

It sits under More in the sidebar, it has been filling in on its own since September 22, 2026, and every employee in your organization reads and writes the same record. What one employee learns about a person, every employee knows, and you can see all of it.

KEY POINTS

7 · 6 min full read

• Contacts, under More in the sidebar, is one record of every person or outside AI agent your AI employees have emailed, called or met. Every employee reads the same record and adds to it, and you see it too.
• It fills in on its own. An email in or out, a call that ends, a meeting that ends and a lead who replies each write the person and the touch. Nobody has to type a record into existence.
• Before an employee answers someone it does not recognise, it looks them up. The answer opens with the relationship: you, a member of your organization, a contact since a given date with a count of exchanges, or no prior contact.
• A vouch is a name on the line. When you or an employee adds someone you have actually dealt with, the record shows who stands behind them, when, and the note they left.
• Notes are for what a colleague would want to know before the first exchange: who introduced them, what they care about, what was agreed. Notes are kept in full and never edited.
• Block is organization wide and for abuse only: a phisher, an impersonator, harassment. Every employee honours it, and you can undo it in one click. Delete forgets a wrong entry or a person who asked to be forgotten.
• Members of your organization are never contacts, only your organization sees its record, and there is no credibility score. The record holds facts; whether to engage stays each employee’s judgment.

§ 01 Why an organization needs one record

A single assistant can keep its own address book. An organization of AI employees cannot. Your receptionist takes a call from a supplier on Monday on its own phone number, your sales employee gets an email from the same person on Wednesday, and your executive assistant meets them on a call on Friday. Without a shared record, each of those employees meets a stranger. With Contacts, the second and third already know who the person is, how your organization came to know them, and what was said.

It is the same thing a well-run office does with a shared contact list, kept up by the people who actually talk to the contacts.

§ 02 Where contacts come from

Nothing needs to be typed for a record to exist. Contacts fill in as your employees work:

• Someone outside the organization emails an employee: The sender, with an email received under that employee
• An employee emails someone outside the organization: The recipient, with an email sent under that employee
• A call to or from an employee’s number ends: The other party’s number, with the call under that employee
• A meeting an employee attended ends: The organizer and each outside attendee, with the meeting
• A lead replies, or was imported with a prior relationship: A contact linked to the lead
• You or an employee add someone by hand: A contact vouched by whoever added them

A first cold email records nothing: the person stays a lead until they answer. That keeps the record to people your organization actually has a relationship with.

§ 03 What your employees do with it

Your employees use the record through five actions of their own. They look someone up before deciding how to handle an email, a call or a meeting invite from a person they do not recognise, and the answer opens with the relationship: you, a member of your organization, a contact since a given date with a count of exchanges and the employees who had them, or no prior contact. They add a person when an exchange is worth remembering or when you introduce someone in chat, with a one-line note on who they are. They list contacts by employee, channel or kind. They block for abuse. They delete a wrong entry.

Employees keep the record the way a person keeps the contacts on their phone: before engaging someone new, they look the person up and, if the exchange is real, save them with a line on who they are. At the end of each working stretch they fill in that line for the people they dealt with, while they still remember them.

None of this decides for the employee. The record says who a person is and how your organization knows them; whether to engage is the employee’s own judgment, the same as a colleague reading the same facts.

§ 04 Vouches and notes

A vouch is a name on the line. When you add a contact, you are saying “I know who this is.” When an employee adds a person it was introduced to or has dealt with directly, it can vouch the same way, with a note saying how it knows them. Employees are asked to vouch only for people they have actually dealt with, never for someone they only read about, and every vouch shows its date and note.

Notes are what a colleague would want to know before a first exchange: who introduced them, what they care about, what was agreed, what to avoid. Anyone in your organization, and any employee, can add one. Notes are kept in full and never edited or deleted.

§ 05 Block and delete

An organization admin can block a contact with a reason, and so can an employee that meets abuse. A block is organization wide: every employee reads it before engaging, on email, calls and meetings. It is meant for a phisher, an impersonator, harassment, or someone asking for data or money under a false name. It is not the tool for “not a good fit”; that stays a judgment each employee makes on each exchange. Every block shows who made it and why, and one click undoes it.

Delete is different. It forgets a person: the contact leaves every list and lookup, and no later email, call or meeting brings them back on its own. It is for a wrong entry or someone who asked to be forgotten, never a substitute for a block.

Your employees already share the record.

Every email, call and meeting your AI employees handle adds to one record your whole organization reads. Hire an employee and it starts with what the others already know. Try it free, no credit card needed.

§ 06 Privacy

• Members are never contacts. Your address, your colleagues on your company domain and your AI employees are filtered out at every point where a contact could be written.
• Only your organization sees its contacts. Every member and every AI employee of your organization reads the same record; nobody outside it does.
• The record holds facts, not the exchanges. Emails, call transcripts and meeting records stay with the employee that handled them; Contacts links to them.
• The deliverability tag is about the address. Each email address carries a small tag saying whether mail to it is delivered or bounces. It is never a judgment of the person.
• Deleting your account deletes the record. Contacts belong to your organization and go with it.

§ 07 The honest caveats

• No import yet. Contacts do not come in from a CRM or Google Contacts today; they fill in from your employees’ work and from people added by hand.
• A record, not a filter. No setting today limits your employees to known contacts. The record informs their judgment; it does not replace it.
• Coverage starts with the work. Your record holds the people your employees have actually exchanged with. An organization that hired its first employee yesterday has a short list, and that is the right length.

§ 08 Getting started

Open Contacts under More in the sidebar. You will find everyone your employees have emailed, called or met, most recent first. Use Add contact to introduce your accountant, a partner or a key customer to every employee at once, with a note on who they are. The full reference is the Contacts guide.

FREQUENTLY ASKED

5 questions

Q1 Where do I find Contacts?

Under More in the CellCog sidebar. One row per contact, most recently in touch first, with search by name, email, company or domain and filters by employee, channel, kind and blocked.

Q2 Who counts as a contact?

A person, or another company’s AI agent, outside your organization, with an email address or a phone number. You, colleagues on your company’s email domain and your AI employees are never contacts. A cold lead becomes a contact when they reply, answer a call, attend a meeting, or were imported as an existing customer, partner, inbound request or opt-in.

Q3 Can my employees block someone?

Yes, for abuse only, and they must give a reason. The block applies to every employee in your organization, shows on the Contacts page with who made it and why, and an organization admin can undo it in one click.

Q4 Can I import contacts from my CRM or Google Contacts?

Not yet. Today contacts fill in from your employees’ work, and you or your employees add individual people by hand with Add contact.

Q5 Does Contacts rate people?

No. It shows facts: how long your organization has known someone, how many exchanges, with which employees, who vouched, and what was noted. There is no score.

Anthropic released Claude Opus 5.5 on September 22, 2026. The same day, every CellCog tier that ran Claude Fable 5.1 moved to it. Agent Core and Max, Creative Core and Max, Team Core and Max: six tiers, one new engine. Flash tiers stay on Gemini 3.8 Flash. Nothing changes on your side; any chat or AI employee on one of those six tiers is already running the new model, at the same pricing.

If you followed our Opus 5.5 tracker through its two days as a rumor page, this is the operator’s half of the story.

KEY POINTS

7 · 7 min full read

• Anthropic released Claude Opus 5.5 on September 22, 2026. The same day, CellCog’s Agent Core, Agent Max, Creative Core, Creative Max, Team Core and Team Max tiers moved to it. Flash tiers stay on Gemini 3.8 Flash.
• Claude Fable 5.1 leaves every tier. Anthropic’s own line on the new model is that it performs at the level of Fable 5.1 on most work, and its launch table puts Opus 5.5 ahead of Fable 5.1 on every published row.
• Core and Max keep their difference in reasoning depth on the new engine: Core runs a lighter setting, Max a deeper one; Creative runs deeper still on both tiers. Team Core and Team Max now differ only in how far their research searches reach.
• Opus 5.5 is priced at $4 per million input tokens and $20 output, against Fable 5.1’s $10 and $50, with cache reads at $0.20. Anthropic says it also produces output more than 30 percent faster than Opus 5.
• Your pricing does not change with this update. You pay for the work, not the hire, and the cost depends purely on how much work you assign.
• Claude Opus 5 stays behind Opus 5.5 as the fallback: when Anthropic’s safeguards decline a request, the chat steps to it without anything on your side.
• Nothing to do on your side. Any chat or AI employee on an Agent, Creative or Team tier, Core or Max, is already running Opus 5.5; memory, roles and task boards carried over untouched.

§ 01 What changed

Agent Core and Agent Max keep the difference they have always had: Core is the everyday tier, one agent in a fast loop with you, and Max thinks deeper per turn. Agent Max is the tier that runs CellCog’s own AI employees, so the model underneath our own company changed too.

Creative Core and Creative Max moved with the rest. Creative is the same agent tuned for taste, and it runs the deepest reasoning settings we ship; the model underneath is now Opus 5.5.

Team Core and Team Max are several agents challenging each other’s findings, the mode behind our deep research work. Both run Opus 5.5 at one balanced reasoning setting; the difference between Team Core and Team Max is now how far their research searches reach.

§ 02 What Opus 5.5 is

Anthropic calls Opus 5.5 the first model in its Claude 5.5 family, built for long-running agentic coding and knowledge work, and says it performs at the level of Claude Fable 5.1 on most work while costing 40 percent less to run than Opus 5. Its published table has Opus 5.5 ahead of Fable 5.1 on every row: 66.4 percent against 55.8 on Terminal-Bench 4.0, 1846 against 1735 Elo on GDPval-AA, 81.8 against 80.7 on OSWorld 2.0. Those are Anthropic’s numbers; independent replication is the open question, and our release record tracks it.

Anthropic list prices per million tokens, September 22, 2026:

• Claude Fable 5.1: $10 input, $0.25 cache read, $50 output
• Claude Opus 5: $5 input, $0.50 cache read, $25 output
• Claude Opus 5.5: $4 input, $0.20 cache read, $20 output

Two facts matter more than the benchmarks for anyone running agents. The token price is 40 percent of Fable 5.1’s, and the cache read, the tokens a model re-reads from previous turns, is $0.20 per million. A standing agent re-reads an enormous amount of context all day: its instructions, its memory, its tools, the conversation so far. That is the row that decides what long-running work costs, and Anthropic says the new model also produces output more than 30 percent faster than Opus 5.

§ 03 Why day one was safe

We have a rule for model upgrades: switch on day one when the successor is a clean step, and wait when it is not. Anthropic’s model documentation lists the same 1M-token context window, the same 128K output limit and always-on adaptive thinking, so every tier’s operating point carries over with only the model name changing underneath. Anthropic’s own read is that Opus 5.5 matches or beats Fable 5.1 on the work it published at 40 percent of the price. Waiting would have cost users capability, and cost us money, for nothing.

Claude Opus 5 stays behind Opus 5.5 as the fallback. When Anthropic’s safeguards decline a request on the new model, the chat steps to Opus 5 without anything on your side, and Anthropic credits the retry. We did the same kind of day-one move with Fable 5.1 on September 1 and with Gemini 3.8 Flash on September 2.

§ 04 What it means for AI employees

An AI employee’s job is to hold intent across a working day: read what arrived, decide what matters, do the work, write down what it learned so the next session picks up as the same mind. Anthropic’s notes on Opus 5.5 point at the traits that matter for that: it puts the most important information up front, writes more plainly, takes fewer steps and fewer tokens per task, and on Anthropic’s own alignment audit is much less likely than recent models to take hard-to-reverse actions or act outside the boundaries it was given. Those are the failure modes an employee running unattended cannot afford.

Whether it delivers is a question for the coming weeks, not launch day. What we can say today is that the model swapped under the work with nothing lost: every employee’s memory, roles, task board and org structure carried straight over, because none of that lives in the model.

§ 05 What does not change

Your pricing. You pay for the work, not the hire, and the cost depends purely on how much work you assign; the details are on pricing. The cheaper model lowers what it costs us to run long conversations; it does not change what you pay today.

Your setup. There is no model picker and no migration step. If a chat or an employee is on an Agent, Creative or Team tier, Core or Max, it is on Opus 5.5.

Your Flash tiers. Agent Flash and Team Flash run exactly what they ran before.

Hire an AI employee and it runs on Opus 5.5 from its first working session, carrying what it learns from one session into the next.

§ 06 What we are watching

Independent benchmark replication of Anthropic’s launch table; how the new model behaves across a full working day of employee work compared with Fable 5.1; and the real cache-read economics on our own bill. Updates land on the Opus 5.5 record, and anything that changes CellCog’s routing lands here.

Your AI employees can now hold their own phone numbers, answer calls on them, and place calls.

Ask an employee in chat for a number and it offers a few in your company’s area code; say yes and the number is live within moments. People who call it talk to the employee: its voice answers within seconds, tells them they are talking with an AI employee at your company, and the employee’s own judgment sits behind every answer, the way it does in a meeting. Once you have verified your business, the employee can also call people who expect to hear from it. Incoming calls have been live for every CellCog employee since September 18, 2026; placed calls, international calls and local numbers outside the United States since September 19.

KEY POINTS

7 · 13 min full read

• Ask in chat and your employee offers a few available numbers in your company’s area code; your yes is the go-ahead and the number is live within moments. It appears on the employee’s profile and in its email signature, and you can forward your existing line to it at your carrier.
• When someone calls, the employee’s voice answers within seconds, the same voice it uses in meetings, on its standing briefing, and tells the caller they are talking with an AI employee at your company. The employee’s own judgment is behind every answer, and every call is transcribed.
• It places calls to people who expect them: an applicant for a job you posted, someone who asked to be called back, a customer, a partner, a person who opted in. Never strangers. Calls go out inside calling hours and daily limits, under your business name.
• Verify your business once on the Phones page with the registration your country issues: an EIN, a Business Number, a Company Number, an ACN, a CIN, a VAT number, a DUNS number, or a trade license. Once it is verified, every employee in your organization can place calls.
• A local number in about fifty countries, on your business’s registration in that country. Calls reach most of the world; the carrier’s charge for an international leg is passed through, and a local number in the person’s country makes those calls domestic.
• The Phone app on each employee’s computer shows recents, the transcript, the employee’s notes and what it did afterwards; the Phones page shows every number, live call, do-not-call entry and calling rule across the organization.
• A number is a recurring monthly charge in credits at the carrier’s price for that country; calls are metered by the time on the call and shown on each call’s record. Release a number any time: callers hear a busy signal for 30 days, then it is gone.

§ 01 The third room

An employee has had a chat and an inbox from the day it was hired. Last week it got a seat in Google Meet and Zoom. The phone is the third room, and it is the same employee in it: the same voice, the same memory, the same task board, the same rule about what it may and may not decide. That was the thing we cared about most. AI phone agents are a mature category and the good ones answer well; what they answer as is a phone agent. A CellCog employee that picks up is the one who also read the applicant’s resume this morning, owns the follow-up task, and will write the email after the call from the same desk.

§ 02 What happens when someone calls

From ring to follow-up:

• Ring: The call reaches the employee’s own number, the one on its profile and in its email signature
• Answer: Its voice picks up within seconds, in the voice it uses in meetings, on its standing briefing, and tells the caller they are talking with an AI employee at your company
• Think: Anything that needs a decision, its memory or a commitment goes to the employee itself while the conversation continues; your Incoming call wake condition decides whether it joins live or reads the full transcript at its next run
• Hand over: Once you have verified your own number, the voice can transfer the caller to you
• Record: The transcript lands in the Phone app with the outcome and the employee’s one-line notes; the audio is not kept
• Follow up: A task, an email or a note, filed where the employee’s other work already lives

One live conversation at a time. An employee that is on the phone, in a meeting or in a chat meeting is busy, so a second caller hears a busy signal and the call shows as missed. The employee sees the missed call and follows up the way a person would, by email if it knows the caller. There is no voicemail and no message taking, because the employee is the one answering; a caller who wants to leave a message just says it, and it is in the transcript with everything else.

§ 03 What it will and will not do on a call

The rule is the one every CellCog employee already follows in email and in meetings: anything binding comes from a decision you or its brief already made, never from the voice on the line.

The authority rule on the phone:

IT DOES:

• Answers questions about your company, its work and the topic, within what it is authorized to say
• Takes the caller’s request and carries it to its task board
• Treats a call from your own verified number as you speaking, the same as a message in chat

IT DOES NOT:

• Commit your company to dates, prices, scope or promises on the strength of a call
• Send, schedule or change anything that needs an approval without confirming with you in chat first
• Treat any other caller’s instruction as an instruction

§ 04 Placing calls: people who expect them

Once your business is verified, one sentence in chat is the go-ahead, in the shape of "call Priya Shah about the PM role; she applied Tuesday; ask her availability for an interview next week." That sentence tells the employee why the person expects the call. It dials from its own number right away, or sets a one-time routine when you name a time. An employee working from a standing task with nobody in the chat needs a standing approval that names the kind of call, for example interview scheduling calls to applicants, up to ten a day.

Who the employee will call:

IT CALLS:

• An applicant for a job you posted
• Someone who asked to be called back
• A customer
• A partner
• A person who opted in to calls from you

IT DOES NOT CALL:

• Strangers, whatever the list
• Anyone on your organization’s do-not-call list, which every employee shares
• Outside 8:00 am to 9:00 pm in the person’s local time (9:00 am to 6:00 pm in the United Arab Emirates)
• Past the number’s daily limit, which is lower while the number is new
• A second time the same day after one retry with no answer

The person hears the same voice the employee uses on every call, and is told at the start that they are speaking with an AI employee at your company. When a call reaches voicemail, the employee leaves one line and hangs up, as a person would; a phone that screens unknown callers can sound like voicemail for a moment, and if the person then picks up and speaks, the employee carries on. When someone asks not to be called again, the employee marks the number, and no employee in your organization calls it again.

§ 05 Verify your business, once

Every call your employees place goes out under your business, never under CellCog: the carriers see your business as the caller, and your legal business name is the caller name where a network shows one. Before the first call, an organization admin verifies the business on the Phones page (under More in the sidebar), under Compliance, Business identity. The form is prefilled with your company name and business address; you add the legal name, your business registration, the business type, the industry and the website, with your own verified phone number as the contact. The carriers review it, usually within one to three days, and once the block reads Verified, every employee in your organization can place calls.

The registration is whatever your country issues, and the form picks the type from your address: an EIN in the United States, a Business Number in Canada, a Company Number in the United Kingdom, an ACN in Australia, a CIN in India, a VAT number in the European Union, a DUNS number from anywhere, or another registration you name, such as a trade license. Two honest details. The phone networks sign calls, so they are not labeled as likely spam, and show a caller name, only for businesses registered with an EIN or a DUNS number; a business verified with any other registration places calls under its own name without that signature, and its status says so and offers to add one (a DUNS number is issued free by Dun and Bradstreet to a company anywhere in the world). And a personal account, one that signed up with a consumer email address, cannot verify a business, so it keeps everything on this page except placing calls.

§ 06 A number where your customers are

An employee can hold a local number in about fifty countries the carrier sells to the platform: the United Kingdom, Canada, Australia, Germany, France, the Netherlands, Japan, Mexico, Brazil and more; the Phones page shows the list. Ask in chat the same way: "get yourself a UK mobile number." Most countries require the business to be registered with the carrier first, with the documents that country asks for, and the Phones page renders that country’s own requirements as a form you fill once; the United States and Canada need no registration. A local number means people in that country reach your employee without dialing abroad, and its calls to them show a local caller ID.

Calls can go to numbers in most of the world; the few countries the carrier does not allow are refused by name, and the employee tells you. The carrier’s charge for an international leg is passed through on top of the call’s usual metering. The United Arab Emirates, India and Singapore are the notable gap on numbers: no vendor sells them to us today, so a business there holds a number from another country and calls home at international pricing.

§ 07 The Phone app and the Phones page

Every employee’s computer carries a Phone app. Until the employee has a number, it is one screen with an Ask button that opens the chat with the request ready to send. Afterwards it looks like the phone app you already know: Recents, one row per call with who, when, how long and the one-line notes the employee wrote; the call detail with the outcome, the transcript, what the employee did afterwards, the credits charged and, for a placed call, why it called and on what basis the person expected it; and Contacts, the people the employee knows who have a number, for reading only.

Once more than one employee has a phone, the Phones page is the one place for all of them: every number in the organization with its state and today’s calls, a live band with End on each call in progress, every call across the organization with filters for missed, employee, received or placed, domestic or international, and outcome, and the Compliance section, where the business identity, the country records, your own verified number, the do-not-call list and the calling rules live. It reads the same records the Phone app does.

An AI employee on CellCog owns a role, keeps a task board and an inbox, sits in your meetings, and now holds a phone number of its own. Ask it for one in chat and read the transcripts in its Phone app afterward.

§ 08 How this compares, from the other platforms’ own docs

Read on September 19, 2026, each from the vendor’s own page. Dedicated AI phone platforms are good at phones; the question we asked each page was narrower: does the agent get its own number, does it place calls, does it tell the caller it is an AI, and is the phone one room of a larger employee or the whole product.

Bland, Retell and Synthflow are dedicated platforms, and they are the right tool when the phone is the product. Read their pages on disclosure, though: it is a guardrail you switch on, a default that fires when the caller asks, or a greeting you may use. Marblism’s Rachel is the closest shape to ours, a named employee with a number of her own, and her page does not say whether she tells callers she is an AI. A CellCog employee opens every call with it, on the calls it takes and the calls it places, and it does not have a disclosure setting to turn off. The other difference is where the phone sits. For the platforms above, the number is the product, with tools around it. For a CellCog employee, the number is one more room it walks into with its memory, its inbox, its meetings and its board, and the same rule about what it may promise.

§ 09 The honest caveats

• Calls only. Texting and a Messages app come next, on the same business verification, so a verified business is ready for them.
• Porting an existing number is not available yet; forward it instead.
• One live conversation per employee at a time; a second caller hears busy and the call shows as missed.
• Businesses verified without an EIN or a DUNS number place unsigned calls, which pick up more likely-spam labels; a free DUNS number fixes it.
• A personal account can hold a number and take calls, and cannot place one.
• No numbers in the United Arab Emirates, India or Singapore today; calls to them, yes.
• Branded calling (your name and logo on the screen of the phone being called) and blocking a number come later.
• How the voice greets, paces and takes turns is its own. The loop is what we can promise: it answers, it is transcribed, the employee’s judgment is behind it, and it follows up.

§ 10 What it costs

A phone number is a recurring monthly charge in credits, the carrier’s own price for that country and number type; the employee quotes the exact figure from its receipt when it offers you numbers, and the charge appears in your credit history each month. Calls are metered by the time on the call, and the employee’s own thinking bills the way it always does; each call’s record shows the credits charged. A call nobody answers costs nothing, and verifying your business is free. The current figures are on the pricing page.

FREQUENTLY ASKED

Q1 Can it take messages or voicemail?
There is no voicemail and no message taking: the employee is the one answering, so a caller who wants to leave a message just says it, and it is in the transcript. When the employee is already on a call, in a meeting or in a chat meeting, a second caller hears a busy signal, the call shows as missed, and the employee follows up the way a person would.

Q2 Can I keep my existing business number?
Porting is not available yet. Forward your existing line to the employee’s number at your carrier, and give the new number out directly wherever you can.

Q3 My business is outside the United States. Can my employee place calls?
Yes. Verify it on the Phones page with the registration your country issues; the form picks the type from your business address. Only an EIN or a DUNS number earns the carriers’ signature and a caller name on US networks; a DUNS number is issued free by Dun and Bradstreet to a company anywhere in the world.

Q4 Can my employee have a number in another country?
Yes, in about fifty countries the carrier sells to the platform, including the United Kingdom, Canada, Australia, Germany, France, the Netherlands, Japan, Mexico and Brazil. Most need the business registered with the carrier first, filled once on the Phones page under Compliance; the United States and Canada need none. The United Arab Emirates, India and Singapore are not sold today; your employee can still call numbers there from another number.

Q5 Is the call recorded?
The transcript is kept with the call record, along with who called, when, how long, the outcome and the employee’s notes; the audio is not kept. Records live with the employee and are deleted when the employee is deleted.

Q6 Can I call my employee myself?
Yes. Verify your own number in the employee’s Settings app, and a call from you is you speaking, the same as a message in chat: the employee treats it as your instruction, and anything that needs your approval still confirms in chat. Once your number is verified, the voice can also transfer a caller to you.

Google approved CellCog’s OAuth app verification on September 13, 2026. The approval covers the seven Google scopes behind two connectors, Gmail and Google Analytics, and it came after a CASA Tier 2 security assessment completed through TAC Security, an App Defense Alliance authorized lab. From now on, connecting Gmail or Google Analytics to a CellCog AI employee shows no unverified-app warning from Google, and the 100-user cap Google places on unverified apps no longer applies to those connects. Nothing changes for connections you already made.

This post is the record of what was assessed, what it changes, and what it does not mean. No email went out for this update; the change is invisible when it works, which is the point.

KEY POINTS

• Google approved CellCog’s OAuth app verification on September 13, 2026, for the Gmail and Google Analytics scopes behind the Gmail and Analytics connects. The gate was a CASA Tier 2 (Assurance Level 1) security assessment by TAC Security, an App Defense Alliance authorized lab.
• What changes for you: connecting Gmail or Google Analytics to a CellCog AI employee no longer shows Google’s unverified-app warning, and the 100-user cap Google places on unverified apps is gone. Existing connections keep working; nobody reconnects.
• Why the bar exists: Gmail’s full-access scope is one of Google’s restricted scopes. Google’s own rule is that apps handling restricted data through a third-party server complete an annual security assessment by a Google-approved assessor. An AI employee that reads and sends your mail is exactly that kind of app.
• The assessment is built on the OWASP Application Security Verification Standard and covers the application itself: how it authenticates, controls access, protects data and handles secrets. The evidence pack CellCog submitted on August 14, 2026 answered 48 checks.
• Plain words, on purpose: this is an assessment, not a certification. Google and the App Defense Alliance issue no badge at this tier; the assessor’s own mark is the only one, shown only as TAC Security delivers it. We say ‘assessment completed’ and ‘Google approved’, dated, and nothing more. It is reassessed every year.
• Scope is narrow and stated: Gmail and Google Analytics on CellCog’s own Google client. Drive, Calendar, Docs, Sheets and the other Google connectors run through a managed integration partner’s app and are outside this assessment.

§ 01 Why an AI employee needs a verified door into Gmail

A chatbot that drafts an email for you to paste never touches your mailbox. An AI employee does. It reads the thread, replies from its own address or works yours, files what it learns, and comes back tomorrow. For that job to be real, the Gmail connection has to be a full one, and Google classifies full Gmail access as a restricted scope. Google’s rule for restricted scopes is direct: “Apps accessing restricted data from or through a third-party server must undergo an annual security assessment by a Google-approved third party.” An organization of AI employees that reads and sends mail on your behalf is exactly the kind of app that sentence describes, so the assessment was never optional for the product we are building. It was the cost of building it honestly.

§ 02 What was assessed

The assessment, from the records:

• What Google approved: OAuth app verification for CellCog’s own Google client, September 13, 2026
• Scopes: Full Gmail (mail.google.com); contacts read; other contacts read; profile email read; Google Analytics read, edit and full access
• The gate: CASA Tier 2 (Assurance Level 1), required by Google’s notice of August 11, 2026
• The assessor: TAC Security, an App Defense Alliance authorized lab
• The evidence: 48 checks against the CASA requirements, submitted on the lab’s portal on August 14, 2026
• The cycle: Reassessed annually

CASA is built on the OWASP Application Security Verification Standard. At Tier 2 the developer tests and documents the application against those requirements and the authorized lab validates the evidence; at Tier 3 the lab tests the application, its deployment infrastructure and its data storage itself. The requirements sit in the categories anyone who has read the ASVS would expect: authentication, session handling, access control, protection of data in transit and at rest, handling of secrets, configuration. What CellCog changed while working through them stays inside the company. The categories are public; the specifics of any one application’s defenses are not something a security-minded company publishes, and we are not going to start.

§ 03 What changes when you connect

Before and after, for the two connectors in scope:

• Connecting Gmail: Before - Google’s unverified-app warning; 100-user cap on the app; Now - Standard Google consent screen; no cap
• Connecting Google Analytics: Before - Same warning, same cap; Now - Same consent screen, no cap
• Existing connections: Before - Working; Now - Unchanged; nobody reconnects
• Drive, Calendar, Docs, Sheets, Search Console, YouTube: Connect through a managed integration partner’s Google app; Unchanged; outside this assessment

The last row matters. This assessment covers Gmail and Google Analytics on CellCog’s own Google client. The other Google connectors run through an integration partner’s app, with that partner’s own verification, and nothing on this page speaks for them.

§ 04 What we say, and what we do not

The line we use, everywhere it appears, is this: Google OAuth app verification approved, September 2026; CASA Tier 2 security assessment completed through an App Defense Alliance authorized lab, reassessed annually. Dated, and specific to the scopes above.

On the marketing pages the same fact reads in plainer words, and the two surfaces use one sentence so nobody has to reconcile them: “Google reviewed how CellCog connects to your Gmail and Google Analytics and approved it in September 2026, after an independent App Defense Alliance authorized lab assessed our application security; both are repeated every year.”

What we do not say: certified. CASA is an assessment, the lab validates evidence, and Google does not audit anyone. Google and the App Defense Alliance issue no badge at this tier; the Alliance’s public disclosure language is for its lab-validated Tier 3 process, and Google’s Workspace Marketplace security badge requires Assurance Level 2. We hold neither and claim neither. The assessor, TAC Security, issues its own mark to the companies it assesses; where that mark appears on cellcog.ai it is shown exactly as TAC delivers it, never redrawn, and never with the word certified beside it. We also do not say SOC 2, ISO 27001, enterprise-grade, or any phrase that borrows the weight of a certification we have not earned.

§ 05 How the other platforms word it, from their own pages

Two companies in our category publish a CASA line, and both are worth reading for what the words carry. Coworker.ai’s Google Workspace page says “Coworker is SOC 2 Type II certified, GDPR compliant, and CASA Tier 2 certified.” 11x’s security overview says “11x holds SOC 2 Type II and CASA Tier 3 certification and complies with GDPR and CCPA.” Both are real assessments and both companies did the work; 11x’s Tier 3 is the deeper of the two, with the lab testing the infrastructure as well as the application, and it is a fair thing to be proud of. The word we would quarrel with is “certified”, which the App Defense Alliance’s own pages do not use for CASA. We passed the same Tier 2 gate Coworker did; we call it an assessment because that is what it is. Buyers who care about this should ask any vendor, us included, for the tier, the assessor, the date and the scopes covered, and read the answer against the App Defense Alliance’s assurance-level page rather than against the adjective.

An AI employee that can read, reply and file mail is the difference between a tool and a colleague. Gmail now connects on Google’s standard consent screen. Try it free, no credit card needed, and you pay for the work, not the hire.

§ 06 What comes next

The assessment renews every twelve months, and the next cycle is on our calendar. The security page on cellcog.ai carries this record and the rest of what we publish about how the product handles your data. The Gmail connection itself needs nothing from you.

§ 07 Sources

Google, Restricted scope verification and Gmail API scopes, read September 16, 2026; App Defense Alliance, CASA overview, assurance levels and self-initiated assessment, read September 16, 2026; Google, Get your app featured in the Google Workspace Marketplace, read September 16, 2026; Coworker.ai, Google Workspace connector page, read September 16, 2026; 11x, Security and compliance overview, read September 16, 2026. Internal records: Google’s approval notice of September 13, 2026 and the assessor’s portal submission of August 14, 2026.

FREQUENTLY ASKED

Q1 What is CASA?
The Cloud Application Security Assessment, run by the App Defense Alliance. It applies the OWASP Application Security Verification Standard to apps that handle certain categories of Google user data, at tiers that differ by how the evidence is tested and verified. Google requires it for apps that access restricted scopes such as full Gmail through a third-party server, and it must be repeated every twelve months.

Q2 What is the difference between Tier 2 and Tier 3?
At Tier 2, now called Assurance Level 1, the developer tests and documents the application against the CASA requirements and an authorized lab validates the evidence. At Tier 3, or Assurance Level 2, the lab itself tests the application, its deployment infrastructure and its data storage. Google’s Workspace Marketplace badge requires AL2; CellCog completed AL1, which is the level Google’s approval required for these scopes.

Q3 Who assessed CellCog?
TAC Security, one of the App Defense Alliance’s authorized labs. The evidence pack was submitted on their portal on August 14, 2026, and Google’s approval arrived on September 13, 2026.

Q4 Which scopes were approved?
Full Gmail access (mail.google.com), read access to contacts and other contacts, read access to the account’s email address, and Google Analytics read, edit and full access. These are the scopes behind CellCog’s Gmail and Google Analytics connects.

Q5 Do I need to reconnect anything?
No. Existing Gmail and Google Analytics connections keep working exactly as before. New connections skip the warning screen.

Q6 What does CellCog not claim?
That it is certified, audited by Google, SOC 2 or ISO 27001 compliant, or that any Google or App Defense Alliance badge applies. CASA is an application security assessment with an annual cycle; that is the claim, dated, and the only one.

NEXT STEP

Give an AI employee your inbox, on Google's terms
Hire one, connect Gmail without the warning screen, and let it work the mail. Roles and honest pricing are on the employees page.

Your AI employees can now attend Google Meet and Zoom meetings, and speak in them.

Add an employee’s email address to a calendar invite the way you would a colleague’s. It accepts, prepares from what it already knows about the attendees and the topic, joins one minute before the start under its own name, listens to the whole room, answers when it is addressed, and after the call writes the notes and files the decisions and action items where its other work already lives. Meeting Presence has been live for every CellCog employee since September 13, 2026.

KEY POINTS

7 · 10 min full read

• You invite an AI employee the way you invite a colleague: add its email to the calendar invite. It accepts or declines by its invite policy, so the organizer’s calendar shows the answer.
• It joins one minute before the start with its camera off, under the name {Name} (AI, {Company}). The name tag says it is an AI, and asked directly it says so plainly.
• It listens to the whole room, knowing who is speaking, and answers when addressed: questions about your company, its work and the topic, within what it is authorized to say. It stops when interrupted and asks its own clarifying questions.
• It never commits your company to dates, prices, scope or promises in the room, and it does not send, schedule or change anything during the call. Anything binding comes back after the meeting, from a decision you or its brief already made.
• You can read the preparation it wrote before the join, whisper it a brief mid-call, mute what it hears, or end the meeting, all from the Meetings app on its computer.
• Several employees can attend one meeting, each under its own name, and a question outside one employee’s lane goes back to the coworker whose lane it is. After the call, the notes, decisions and action items land on the task board.
• Recording is off by default: the employee keeps a transcript and its notes, and the audio is not kept. Attending is metered by the minute it is in the call, and the exact cost in credits shows on the meeting’s record.

§ 01 Why a colleague, not a recorder

Meeting notetakers are everywhere. They join, record, transcribe and send a summary, and that is a fine tool for one person who cannot make a call. An organization of AI employees needs something else. The employee who runs your outbound desk should be able to sit in the weekly pipeline review and answer how many first touches went out last week when the founder asks, the way the human in that seat would. The employee who owns your blog should be able to say which posts are live. Neither should promise a customer a date it has no authority to promise.

So we built the meeting as one more place the employee does its job, with the same memory, the same task board and the same rules about what it may and may not decide.

§ 02 What happens when you invite one

From invite to follow-up:

• Invite: You add its email to the invite in Google Calendar, Zoom or Outlook, or ask it in chat to join a call with the link
• Decide: It checks its invite policy and accepts or declines, so the organizer’s calendar shows the answer
• Prepare: It writes a preparation from what it already knows: who is coming, its email history with each attendee, the goal, what it may say and what to defer. You can read it in the Meetings app before the join
• Join: One minute before the start, camera off, under the name {Name} (AI, {Company}); it waits in the waiting room if there is one
• In the room: Listens to everyone, knowing who is speaking; answers when addressed; stops when interrupted; asks its own clarifying questions; says goodbye when asked or when the room wraps up
• After: Writes the notes, files the decisions and action items on its task board and, when mandated, in email; the transcript and the preparation stay on the meeting record

The name tag says it is an AI, and asked directly it says so plainly. It comes in quietly. In a room of several people it is meant to stay quiet until someone puts a question to it, and it always answers a hello or goodbye said to it by name. If everyone else has left, it leaves too.

§ 03 What it will and will not say

The rule is the one every CellCog employee already follows in email: anything binding comes from a decision you or its brief already made, never from the voice in the room.

The authority rule in a meeting:

IT ANSWERS:

• Questions about your company, its work and the topic, within what it is authorized to say
• What is in its notes and its preparation, almost at once
• A harder question, thought through while the conversation continues and brought in at the next natural gap

IT DEFERS:

• Dates, prices, scope, promises: anything that commits the company
• Sending, scheduling or changing anything during the call
• Anything outside its brief, which goes back to its own judgment and to your standing decisions after the call

When a question needs a decision it does not hold, it says in its own words that it is looking into it and will come back, lets the conversation continue, and either brings the answer at the next gap or follows up after the call. A Never say list in its settings keeps specific lines out of every meeting regardless of what it is asked: pricing beyond your public page, other customers’ names, unannounced work.

§ 04 What you control

Every employee’s computer now has a Meetings app with three tabs.

Upcoming lists accepted invites with the platform, the organizer, the attendee count and a Decline button.

Live shows who is in the room and who is speaking, the transcript as it lands, and three controls:

• Whisper sends the employee a brief mid-call (a line such as: do not quote a price),
• Mute stops it hearing the room until you unmute,
• End makes it say goodbye and leave.

Past holds the transcript, the preparation it wrote, the cost of the meeting and the recording when one exists.

Three settings decide the shape of the whole thing:

• Invite policy. By default the employee accepts invites from you and from people it has already emailed with, and asks you about everyone else. You can widen that to any invite or narrow it to always ask first. Every employee attends by default; switch it off under Meetings in its Settings app and it declines politely.
• Recording. Off by default: the employee keeps a transcript and its notes, and the audio is not kept. Turn it on per employee and it announces the recording within the first minute of every call.
• Voice. The employee chose its own voice when it named itself; you can change it any time in Settings. English by default, and it switches language when someone asks (it has answered in Hindi in a live test).

§ 05 Several employees, one room

Any employee in your organization can be invited, and two can attend the same meeting; each joins under its own name. We tested that on September 14, 2026 with two of our own. They answered the questions put to each of them and talked to each other when asked to. Each employee knows the company picture going in (how many AI employees the company has, who they are, that each has memory, email, internal channels, tools and a task board), so a question outside its own lane goes back to the employee whose lane it is, after the call, instead of to a guess in the room.

This is the part a notetaker cannot do, and it is why we built the feature at the level of the organization rather than the assistant. Whoever invited a CellCog employee to a meeting invited a colleague with a job, a memory and coworkers.

§ 06 How this compares, from the other platforms’ own docs

Read on September 15, 2026, each from the vendor’s own page. Two things are documented widely: joining a call as a visible participant, and recording it. A participant that answers in the room under authority rules is what we could not find.

Fireflies is the closest thing to a talking participant on that list, and it is a good product: a notetaker you can call on by name for a fact from the discussion or a bookmark. The difference is where the voice sits. A CellCog employee is not a tool the room queries; it is a colleague with a role, a memory of every email it has exchanged with the people on the call, a task board that receives the action items, and a rule about what it is not allowed to promise. Grok Bot’s meeting stand-in is the other design worth watching: xAI describes it as joining when you cannot make it and telling the room you are there, and we will fold in what its docs say about speaking when they say it.

§ 07 The honest caveats

• Microsoft Teams is not supported yet; it comes after Meet and Zoom.
• Camera off, presence tile only: no video presence yet, and it cannot present a screen or share a document.
• Zoom meetings that require an authenticated Zoom account cannot be joined, and waiting rooms on either platform need the host to admit it. If it could not get in, the Past tab says so and the employee tells you.
• Group etiquette is the employee’s own judgment, not a rule. It may greet or acknowledge at a moment a human would not. Mute and End are there for that.
• It never commits or acts inside the call; anything binding comes after it.

§ 08 What a meeting costs

Attending is metered by the minute the employee is in the call, and the employee’s own thinking bills the way it always does. Each past meeting shows its exact cost in credits on its record in the Meetings app, the amount deducted from your balance for that meeting. The current figures are on the pricing page.

FREQUENTLY ASKED

Q1 Which meeting platforms work today?

Google Meet and Zoom. On Meet the employee joins as a guest and asks to join; if the meeting admits invitees only, the host admits it from the waiting room. On Zoom it joins like a participant; meetings that require an authenticated Zoom account cannot be joined. Microsoft Teams comes after Meet and Zoom.

Q2 Who can invite my employee?

By default it accepts invites from you and from people it has already emailed with, and turns any other invite into a question for you with one-click Accept or Decline. You can widen that to any invite or narrow it to always ask first. Every employee attends by default; switch it off under Meetings in its Settings app and it declines politely.

Q3 Is the meeting recorded?

Not unless you turn recording on for that employee. Off by default, it keeps a transcript with speaker names and its own notes; the audio is not kept, and the voice service it speaks through keeps no recording either. With recording on, it announces the recording within the first minute of the call and the file lands next to the transcript.

Q4 What can I do while the meeting is running?

The Live tab of the Meetings app shows who is in the room, who is speaking and the transcript as it lands, with three controls: Whisper sends the employee a brief mid-call, Mute stops it hearing the room until you unmute, and End makes it say goodbye and leave. Whisper also works before the meeting, from the Upcoming tab, and reaches the employee while it prepares.

Q5 Can two of my employees attend the same meeting?

Yes, if both are invited. Each joins under its own name, each knows the company picture going in, and they can talk to each other when asked to. We tested it with two of our own employees on September 14, 2026.

Q6 What language does it speak?

English by default, and it switches when someone asks; it has answered in Hindi in a live test. You choose its voice in the Settings app on its computer, and the voice stays fixed for the length of a meeting.

NEXT STEP

Put an AI employee in your next meeting

Hire one, add its email to the invite, and read the notes on its Meetings app afterward. Roles and honest pricing are on the employees page.

Every scheduled wake across your organization now lives on one page. A routine is a named time trigger with a standing brief: it wakes one AI employee, a whole team, or a workstream at the moments you choose, on a full recurrence rule, and the new Routines page shows all of them together: who works, when, and with what mandate. Beside it, every employee’s computer now has a Wake app, which gathers everything that can bring that employee online, from a routine to an inbound email to an event in one of your connected tools, each with its own wake policy.

An organization runs on time as much as on tasks. A single assistant gets by with “every weekday at 8.” A company of AI employees needs a support desk that opens at 9 in three timezones, a finance close that runs on the 1st, a market watch that checks in several times a day, a team that wakes together for a synchronized working session, and a manager’s view that shows load, gaps and one-time reminders across everyone at once. Routines were built for the second case.

We run CellCog on routines ourselves, and the design below comes out of that use.

KEY POINTS

7 · 13 min full read

• A routine is a named time trigger with a standing brief. It can wake one AI employee, every employee on a team, or every employee on a workstream, and the brief is injected fresh at every fire.
• The schedule is one full recurrence rule (RFC 5545): weekdays at 9am, three times a day, the 1st of the month at 7am, or once next Tuesday. How often it runs lives in the rule itself, so a one-time reminder and a forever schedule are the same object.
• Every wake carries a policy: Always, If eco (wake only while the previous session’s context is still cached) or Next start (never wake; the brief rides into the next session). No brief is ever lost.
• A fire that lands while the employee is already working merges into the live session. Two routines in the same minute, or a routine plus you chatting, still produce one session.
• The Routines page lists every scheduled wake across your organization: who works, when, with what brief, plotted on a week strip with per-day run counts. Employees write their own routines from chat when your ask carries a clock.
• Every employee’s computer has a Wake app: routines plus the built-in conditions (your message, an email, a task comment, a coworker’s delegation, a channel mention) and events from your connected tools, each with the same three policies.
• Of the six platforms whose scheduling docs we read on September 4, 2026, none publishes an org-wide schedule view, and only Relevance AI documents one schedule that wakes several agents. Receipts and links below.

§ 01 What a routine is

Six fields, and the schedule is one of them.

The anatomy of a routine

FIELD - WHAT IT HOLDS
Name - “Morning support desk”, “Weekly pipeline review”, “Month-end close”
Schedule - One recurrence rule in the RFC 5545 grammar, read in the routine’s timezone
Brief - The standing mandate, up to 4,000 characters, injected fresh at every fire; edit it and every future fire changes
Assigned to - One employee, one team, or one workstream
When it fires - Always, If eco, or Next start (below)
Ends - Never, on a date, or after N times; the rule carries it as UNTIL or COUNT

How often a routine runs is not a type you pick; it lives in the rule.
FREQ=DAILY;BYHOUR=6,12,18;BYMINUTE=0 wakes an employee three times a day.
FREQ=MONTHLY;BYMONTHDAY=1;BYHOUR=7;BYMINUTE=0 runs once a month, on the 1st. A one-off reminder is a rule with COUNT=1: it fires once, marks itself completed, and drops out of the beat and the default lists. The form hides the grammar when you do not need it (Once, Repeats on these days at this time, Custom for everything else), and the schedule is always shown back to you in words with its timezone, so “Once, Tue Sep 8 at 7:30 AM ET” never reads as “Daily”. The full grammar we honor is the iCalendar recurrence rule: FREQ from minutely to yearly, INTERVAL, BYDAY, BYHOUR with several values, BYMINUTE, BYMONTHDAY, COUNT, UNTIL, and a DTSTART line with a TZID.

§ 02 Three kinds of target: employee, team, workstream

A routine assigned to an employee wakes that employee. Assigned to a team, it wakes every AI employee on the team (the team’s AI leader included) in the same window, so they can work a shared problem live over channels and DMs. Assigned to a workstream, it wakes every employee on that line of work for scheduled progress on the thing they share.

Membership resolves at fire time. Add someone to a team on Tuesday and Wednesday’s routine wakes them; the routine itself never needs editing. A shared routine has a Members panel where each member has its own participation switch and can carry its own wake policy, so one employee sits out a team routine without touching it for the others. Each member fires through its own gates and bills its own owner: assigning a routine can never spend someone else’s credits. And a paused employee never fires, whatever its routines say.

This is the piece that turns a scheduler into an organizational tool. The org chart says who reports to whom; routines say when the team is in the room together.

Hire the first employee, then give it a team.

An AI employee on CellCog owns a role, keeps a task board and an inbox, and wakes on the clock you set. Hire two and put them on a team routine, and they work the same window together.

§ 03 Wake policies: Always, If eco, Next start

Every wake, whether it comes from a routine or from an inbound email, carries one of three policies.

• Always. Wake for every fire. If the employee is mid-session, the brief merges into the live session as a note it reads on its next turn. If the last session has concluded, the employee resumes in place when its context is still cached, or a fresh session starts with the brief already in its mailbox.
• If eco. Wake only while resuming the previous session is still eco-priced, meaning its context is still cached. Otherwise the brief is held and waits for the next run that happens anyway; no session is ever started for it.
• Next start. Never wake. The brief rides into whatever session runs next.

No brief is ever lost in the quieter two modes; it waits in the employee’s mailbox. Which means a routine can carry a standing instruction (“when you are next up, refresh the churn board”) without paying for a wake, and an org can set its own balance between responsiveness and spend, per routine and per member.

§ 04 What happens at a fire, and when two land at once

The scheduler scans for due routines every minute. At a fire, the routine’s next run is advanced first, atomically, and only then is each member touched: participation switch, pause state, wake policy, credit balance, delivery. One ledger entry per fire records each member’s outcome (started, merged, held for next start, gated on credits, paused), and the routine keeps its last twenty.

Because the advance runs before dispatch and delivery re-checks the session’s state under a single lock, a fire that lands while the employee is already working merges into that session instead of starting a second one. Two routines due in the same minute, or a routine landing while you are chatting with the employee, produce one session with both briefs in it. Duplicate wakes are the classic failure of agent schedulers; this design rules them out.

Missed fires are owed once. Because the next run is always “the next occurrence strictly after now,” a routine paused for a week, or one whose owner could not fund runs for three days, comes back due one time, not seven.

Run now is a real test fire through the same pipeline (members wake, the ledger records it) that moves neither the schedule nor the run counter.

§ 05 Routines and recurring tasks are different things

A routine wakes an employee at a time. A recurring task on the task board is a to-do the employee handles whenever it is awake. The razor: work that must happen at a specific moment (a deadline, a delivery window, a synchronized team session) is a routine, because being woken at the right time is part of doing it right. Work that just needs to happen roughly each working day is a recurring task.

Recurring tasks never wake anyone, so every cadence on a board needs a routine that could host it. The platform checks this mechanically: when an employee creates a task with a cadence (or changes one), a coverage analyzer projects the task’s due dates and the employee’s routine fires over 28 days and returns a verdict, covered, loose or uncovered, along with a ready-to-create routine when the cadence has no clock. Pausing or deleting a routine warns when it is the only clock rescuing tasks. The employee’s three honest answers to a gap: add a hosting routine, coarsen the cadence to match a clock it already has, or consciously accept best effort.

§ 06 Who writes the routines

You do, from the Routines page or an employee’s Wake app. The employee does, from chat: if your ask carries a clock (“send me the report before 8am”, “email them Thursday at 3pm”), creating the routine is part of doing the task, and it tells you the name, schedule and next fire. If the employee invents a clock itself (a responsibility with no dedicated time, found during its own reflection), it proposes first, so repetition never silently turns into standing spend. A coworker employee can create a routine for a teammate, a team or a workstream it can reach, the same way it delegates a task.

You see, edit and can pause every routine in the organization, whoever created it. An employee can delete only the routines it created; owner-created routines it can at most pause. Anyone a routine targets is notified when it is created, edited, paused or deleted.

§ 07 The Routines page and the Wake app

The Routines page is the organization view: All, Active and Paused counts; search; a per-employee filter that resolves direct, team and workstream targeting on the server; a week strip with per-day run counts and a Repeats / Runs once / Paused legend; rows grouped Today, This week, Later, Paused and Done; a Team or Workstream chip, a Once chip for single-run rules, the schedule in words with its timezone, the relative next fire, a Run now button and the pause switch. Completed routines are hidden by default and one click away.

The Wake app lives on each employee’s computer and answers one question: what can bring this employee online? Three sections, in order.
Built-in: a message from you, an email to the employee, a new task comment, a task assigned by a coworker, and channel activity (a mention or a channel-wide ping). Each has the same three-policy control, an on/off switch and a seven-day fire sparkline; built-ins cannot be deleted, only switched off or made quieter.
Routines: the employee’s rows from the org page, with the week view and the same controls.
Connected: events from your connected tools. Pick a connected toolkit, pick one of its triggers, pick the delivery mode, and that event wakes the employee (or holds the payload for its next start) with the same three policies. Time lives in routines; events live in Connected; one app shows both.

§ 08 How this compares, from the other platforms’ own docs

Scheduling is not new; every serious agent product has some form of it. What differs is the unit being scheduled and whether anyone can see the whole organization’s clock. Here is what six platforms’ published documentation says, read on September 4, 2026 (links go to the pages that say it).

Scheduling features as documented by each platform, September 4, 2026

PLATFORM - SCHEDULE GRAMMAR - WHAT ONE SCHEDULE WAKES - ORG-WIDE SCHEDULE VIEW - WAKE POLICIES
CellCog routines - Full RFC 5545 recurrence rule - One employee, a team, or a workstream - Yes, the Routines page - Always / If eco / Next start
Grok Bot - Natural language (“Every weekday at 8:00 AM”) - One Bot (“tells one Bot when to run”); 50 routines per Bot - Not documented - Not documented
Lindy - Natural language (“Every weekday at 9”); daily, weekly or monthly - One Lindy; workspace routines post to a Slack channel or thread - Routines page with Personal and Workspace tabs - Not documented
Relevance AI - Six-field cron (EventBridge syntax) - One agent, or several agents in a Workforce - Not documented (per agent, per Workforce) - Not documented
Coworker.ai - Hourly, daily, weekly - One agent - Not documented - Not documented
Manus - Natural language, custom patterns - One task - Not documented; limits depend on plan - Not documented
OpenClaw - cron, “every”, “at”, plus a heartbeat - One agent or session - Not documented - Heartbeat defers when the agent is busy

Two things stood out in that reading. First, none of the six publishes a view of every scheduled run across an organization; schedules are found per agent, per Bot or per Workforce. Second, none documents a wake policy: a schedule either wakes the agent or it is off. Both gaps are what you notice the moment an organization has more than a handful of AI employees, because the question stops being “when does my assistant run?” and becomes “who is working at 3am, on whose credits, and did the Tuesday window overlap with the Monday one?” That is the question the Routines page answers, and the wake policies are how an organization keeps two dozen clocks from turning into two dozen bills. If you are weighing platforms on this axis, read the category comparison alongside the vendors’ own pages; where a competitor’s docs improve, we update the table.

§ 09 The honest caveats

Some of this is still coarse. There is no timed pause yet (“skip the next two fires” is today a pause plus a one-run reactivation routine). The week strip shows the current week only; there is no navigation to next week. Run now does not count toward the fire counter, on purpose, and the ledger keeps the last 20 runs per routine rather than a full history. A fire lands within a minute of its slot, not on the second. Twelve live routines per target is the cap. And events and time are separate objects: a routine is time-only, and an event trigger is a wake condition in Connected, not a routine, so “event-triggered routine” is not a phrase we use.

What is true today: one recurrence grammar for a one-off reminder and a forever schedule, three targets including a whole team, three wake policies that let a brief wait without a wake, a fire that merges instead of duplicating, missed runs owed once, and one page for the entire organization’s clock. If you have been running a company of AI employees on a pile of per-agent cron jobs, this is the layer above them.

FREQUENTLY ASKED

6 questions

Q1 What is the difference between a routine and a recurring task?
A routine wakes an employee at a time. A recurring task is a to-do the employee handles whenever it is awake. Work that must happen at a specific moment (a deadline, a delivery window, a synchronized team session) is a routine; work that just needs to happen roughly each working day is a recurring task. Recurring tasks never wake anyone, so every cadence on a task board needs a routine that could host it, and the platform tells the employee when one does not.

Q2 What recurrence rules does a CellCog routine accept?
Any rule in the RFC 5545 recurrence grammar: FREQ (minutely through yearly), INTERVAL, BYDAY, BYHOUR with several values, BYMINUTE, BYMONTHDAY, COUNT, UNTIL, and a DTSTART line with a timezone. The form offers Once, Repeats (days plus a time) and Custom (any rule), and an Ends control (never, on a date, after N times). A routine that ends after one run marks itself completed when that run has happened.

Q3 What do Always, If eco and Next start mean?
They are the three wake policies, and they apply to routines and to every other wake condition. Always: wake for every fire; a live session receives the brief in place, a concluded one restarts. If eco: wake only while the previous session’s context is still cached, so the wake is eco-priced; otherwise the brief waits for the next run that happens anyway. Next start: never wake; the brief is read at the next session. The quieter two attach a schedule to a brief without paying for a wake.

Q4 What happens if a routine fires while my employee is already working?
The brief is delivered into the running session and the employee picks it up as part of its work. No second session starts. The same holds when two routines fire in the same minute or when you are chatting with the employee at the moment a routine fires.

Q5 Who can create routines?
You, from the Routines page or an employee’s Wake app. The employee itself, from chat: if your ask carries a clock (a daily report before 8am, an email on Thursday at 3pm), creating the routine is part of doing the task, and it tells you the name, schedule and next fire. A coworker employee, for a teammate, a team or a workstream it can reach, the way it delegates a task. You see, edit and can pause every routine, whoever created it; an employee can delete only the routines it created.

Q6 What are the limits?
Twelve live routines per target (per employee, per team, per workstream). The scheduler checks for due routines every minute, so a fire lands within a minute of its slot. Each routine keeps its last 20 runs in a ledger. A paused employee never fires, whatever its routines say, and a routine paused for a week comes back owing one run, not seven.

NEXT STEP

Give your AI employees a clock

Hire an AI employee, tell it what should happen at 8am on weekdays, and watch the routine appear on the Routines page with its brief and next fire. Roles and honest pricing are on the employees page.

KEY POINTS

7 · 7 min full read

• Google released Gemini 3.8 Flash on September 2, 2026. The same day, CellCog’s two Flash tiers, Agent Flash and Team Flash, moved from Gemini 3.7 Flash to it.
• The operating point did not change: the same medium thinking setting, the same 1M-token context window, the same 64K output limit. Only the model id moved.
• Google held the price: $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026, exactly what 3.7 Flash cost. Your pricing does not change either.
• Google’s own table puts 3.8 Flash ahead of 3.7 Flash on every published row, led by DeepSWE v1.1 (73.7% vs 65.3%), OSWorld-2.0 (59.0% vs 50.6%), and Terminal-bench 2.1 (89.4% vs 85.8%). Vendor numbers, labeled as such.
• The honest caveat, in Google’s words: 3.8 Flash ‘works harder’ and may use more tokens on complex tasks. Same per-token price, possibly more tokens per job. We kept the medium setting for exactly that reason and are measuring Flash cost per turn over the next week.
• Gemini 3.7 Flash stays in our catalog as a one-line rollback. Core, Max, and Creative tiers are unchanged.
• Nothing to do on your side: any chat or AI employee on a Flash tier is already running Gemini 3.8 Flash.

§ 01 What changed

Two of CellCog’s tiers moved. Five did not.

Agent Flash and Team Flash moved from Gemini 3.7 Flash to Gemini 3.8 Flash. Everything else about their operating point is untouched: the same medium thinking setting we adopted on August 13 with 3.7 Flash, the same 1M-token context window, the same 64K output limit. The routing table that picks the model per request now names 3.8 Flash on exactly those two rows; the same model also took over the large-document reading we do behind the scenes.

Agent Core, Agent Max, Team Core, and Team Max run on Claude Fable 5.1, where they moved yesterday. Agent Creative runs on Claude Opus 5. Neither changed today. If you want the full picture of how modes and tiers fit together, it is in Flash, Core, Max.

§ 02 What Gemini 3.8 Flash is

Google’s newest Flash-class model, and by Google’s count its third Flash release in six weeks. It is a drop-in successor to 3.7 Flash on paper: the developer guide lists the same 1,048,576-token context window, the same 65,536-token maximum output, and the same low, medium, and high thinking levels, with medium as the default. Google held the price too: $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026, then $1.50 and $7.50 from January 1, 2027, the same schedule 3.7 Flash already had.

What moved is the benchmark column, most of all on the agent-shaped rows. These are Google’s numbers from the announcement, and we treat them that way.

Google’s published numbers: Gemini 3.8 Flash vs 3.7 Flash (September 2, 2026)

• Terminal-bench 2.1 (agentic terminal coding): 89.4% vs 85.8%
• DeepSWE v1.1 (long-horizon software engineering): 73.7% vs 65.3%
• Vals Finance Agent v2: 61.4% vs 59.0%
• OSWorld-2.0 (agentic computer use): 59.0% vs 50.6%
• HLE-Verified (expert reasoning): 54.9% vs 53.6%
• Terminal-bench 4.0 (general agent capabilities): 19.1% vs 11.2%
• Gray Swan prompt injection, attack success within 15 attempts (lower is better): 5.5% vs 9.2%

The last row in the table matters most to us. A Flash tier reads a lot of untrusted text: web pages, inbound email, documents somebody else wrote. Google charts 3.8 Flash at a 5.5% attack success rate on Gray Swan’s indirect prompt-injection benchmark, down from 9.2% for 3.7 Flash and beside the Claude line. That does not close the problem we described in our prompt-injection post, but it moves the arithmetic in the right direction for exactly the class of work the fast lane does.

§ 03 Why day one was safe

We have a rule for model upgrades: switch on day one when the successor is a clean superset, and wait when it is not. Gemini 3.8 Flash is a clean superset. Same limits, same effort dial, same API surface, same price. We ran it on our infrastructure before the switch, and Gemini 3.7 Flash stays fully wired as a one-line rollback; Google announced no shutdown date for it.

We did the same with 3.7 Flash on the day Google shipped it, and with Fable 5.1 yesterday. The reasoning is the same each time: when waiting buys no safety, it only costs users capability.

§ 04 The honest caveat

Google says the gains come from a design choice: 3.8 Flash “works harder,” running extra reasoning steps and calling tools iteratively on complex tasks, and “might use more tokens to maximize performance, especially at higher effort levels.” A bill is price times tokens. Same per-token price, possibly more tokens per job.

Two decisions follow from that sentence. We kept the medium thinking setting rather than raising it, because Google names medium as the level that still solves complex agentic tasks while limiting token use, and a higher setting would compound the increase on the cheap tier. And we are measuring, not estimating: Flash cost per turn and per shift, the week before the switch against the week after. If the number climbs more than we like, the levers are a lower thinking setting or 3.7 Flash, both one line away. Your pricing does not move in any of those cases.

§ 05 What does not change

Your pricing. You pay for the work, not the hire, and the cost depends purely on how much work you assign.

Your setup. There is no model picker and no migration step. If a chat or an employee is on a Flash tier, it is on Gemini 3.8 Flash.

Your other tiers. Core, Max, and Creative run exactly what they ran before.

The model changed. The employee didn't.

Hire an AI employee and its fast lane runs on Gemini 3.8 Flash from its first shift, carrying what it learns from one shift into the next.

§ 06 What we are watching

Three things over the next few weeks: Flash token use and cost per turn against the pre-switch baseline, any change in refusal behavior on the Gemini lane under Google’s stricter safety posture, and independent replication of Google’s benchmarks. Updates land on the launch post, and anything that changes CellCog’s routing lands here.

FREQUENTLY ASKED

5 questions

Q1 Which CellCog tiers run on Gemini 3.8 Flash now?

Two: Agent Flash and Team Flash. Both moved from Gemini 3.7 Flash on September 2, 2026 with their operating point otherwise untouched, including the medium thinking setting we have run since August 13. Agent Core, Agent Max, Team Core, and Team Max run on Claude Fable 5.1; Agent Creative runs on Claude Opus 5. How the tiers fit together is in Flash, Core, Max.

Q2 Why did you switch on day one?

Because Gemini 3.8 Flash is a drop-in successor. Google’s developer guide lists the same 1,048,576-token context window, the same 65,536-token output limit, and the same low, medium, and high thinking levels as 3.7 Flash, at the same price. We verified it on our infrastructure before switching, and 3.7 Flash remains one setting away as a fallback. We did the same with 3.7 Flash on August 13 and with Fable 5.1 on September 1: when the successor is a clean superset, waiting costs users capability for nothing.

Q3 Will Flash work cost more now?

Your pricing is unchanged. On our side, the per-token price is identical to 3.7 Flash, but Google states that 3.8 Flash executes extra reasoning steps and tool calls on complex tasks and may use more tokens, especially at higher effort levels. That is why we kept the medium setting rather than raising it, and why we are measuring Flash cost per turn over the first seven days. If it climbs more than we like, the levers are a lower thinking setting or 3.7 Flash, both one line away.

Q4 How much better is Gemini 3.8 Flash than 3.7 Flash?

On Google’s announcement table, every published row moved up: DeepSWE v1.1 73.7% vs 65.3%, OSWorld-2.0 59.0% vs 50.6%, Terminal-bench 2.1 89.4% vs 85.8%, Vals Finance Agent v2 61.4% vs 59.0%, HLE-Verified 54.9% vs 53.6%. Google also charts both 3.8 models at a 5.5% prompt-injection attack success rate on Gray Swan’s benchmark, down from 9.2% for 3.7 Flash. These are Google’s runs; independent replication is the open question, and we will note what holds up.

Q5 What is the difference between Flash and Core now?

The model class. Flash tiers run a Flash-class model for fast, high-volume work where speed and cost matter more than maximum reasoning depth; Core and Max run Claude Fable 5.1 at a lighter and a deeper reasoning setting. Agent tiers are one agent working in a loop with you or running an AI employee’s shift; Team tiers are several agents challenging each other’s findings.

ALREADY LIVE

Hire the employee, not the model

The model swapped under the work; the memory, roles, and org structure stayed. Hire an AI employee and its fast lane runs on Gemini 3.8 Flash from its first shift.

KEY POINTS

7 · 7 min full read

• Anthropic released Claude Fable 5.1 on September 1, 2026. The same day, CellCog’s Agent Core, Agent Max, Team Core, and Team Max tiers moved to it.
• Agent Max keeps its exact operating point: the model swapped from Fable 5 to Fable 5.1 with the same thinking effort, the same 1M-token context window, and the same 128K output limit. The other three tiers moved to Fable 5.1 from the Claude Opus line.
• Core and Max now share one engine and differ in reasoning depth: the Core tiers run Fable 5.1 at a lighter reasoning setting, the Max tiers at a deeper one.
• Fable 5.1 is Anthropic’s most capable generally available model, positioned for long-running agent work. Anthropic reports 55.8% on Terminal-Bench 4.0 against 42.0% for Fable 5.
• Anthropic held Fable 5.1’s headline price at Fable 5’s $10 and $50 per million tokens and cut cache reads by 75%. Standing agents re-read their context every turn, so that cut lands squarely on the kind of work CellCog runs.
• Your pricing does not change with this update. Flash tiers still run on Gemini Flash (3.8 Flash since September 2, covered in its own update) and Agent Creative still runs on Claude Opus 5.
• Nothing to do on your side: any chat or AI employee on an Agent or Team tier, Core or Max, is already on Fable 5.1.

§ 01 What changed

Four of CellCog’s tiers moved. Two did not.

Agent Max moved from Fable 5 to Fable 5.1 with the rest of its operating point untouched: the same thinking effort, the same 1M-token context window, the same 128K output limit. This is the tier that has run CellCog’s own AI employees since June, so the model underneath our own company changed too.

Agent Core moved from the Claude Opus line to Fable 5.1 at a lighter reasoning setting. Agent Core is the everyday tier: one agent in a fast loop with you, balanced for speed and cost.

Team Core and Team Max moved from the Claude Opus line to Fable 5.1. Team tiers are several agents challenging each other’s findings, the mode behind our deep research work. Team Core runs the lighter reasoning setting; Team Max runs deeper.

That leaves Core and Max sharing one engine and differing in one thing: how hard the model thinks per turn. Everything else stays where it was. Flash tiers run on Gemini Flash, 3.8 Flash since September 2. Agent Creative runs on Claude Opus 5.

§ 02 What Fable 5.1 is

Anthropic describes Fable 5.1 as a Mythos-level model built for ambitious, long-running projects, one that avoids easy-seeming shortcuts, fixes the root causes of problems rather than the symptoms, and keeps you updated as it works. Its published numbers lead with terminal work: 55.8% on Terminal-Bench 4.0 against 42.0% for Fable 5, and 52.6% on Terminal-Bench-Science 0.1. It shipped alongside Mythos 5.1, a restricted variant for vetted cyberdefense and life-science organizations that most builders will never touch.

Two facts matter more than the benchmarks for anyone running agents. Anthropic held the headline price at Fable 5’s $10 per million input tokens and $50 per million output tokens. And it cut cache reads by 75%, to $0.25 per million tokens. Cache reads are the tokens a model re-reads from previous turns, and a standing agent re-reads an enormous amount of context all day: its instructions, its memory, its tools, the conversation so far. Anthropic estimates the cut saves about 25% on typical workloads and up to about 45% on highly agentic ones. CellCog’s workloads sit at the agentic end of that range.

The confirmed spec record and our grading of the rumor chain live on the Fable 5.1 tracker.

§ 03 Why day one was safe

We have a rule for model upgrades: switch on day one when the successor is a clean superset, and wait when it is not. Fable 5.1 is a clean superset. Anthropic’s model documentation lists the same 1M-token context window, the same 128K output limit, and the same always-on adaptive thinking as Fable 5. That means Agent Max’s proven operating point carries over with just the model name changing underneath.

The three tiers coming from the Opus line were a separate call, and the same evidence made it. Fable 5.1 at a lighter reasoning setting is a step up from what those tiers ran before, at prices that hold once you account for how much of a long conversation is cached context. Claude Opus 5 stays in our catalog as the fallback behind it.

We did the same with Gemini 3.7 Flash on the day Google shipped it, and with Fable 5 itself in June. The reasoning is the same each time: when waiting buys no safety, it only costs users capability.

§ 04 What it means for AI employees

An AI employee’s job is to hold intent across a shift: read what arrived, decide what matters, do the work, write down what it learned so the next shift picks up as the same mind. Fable 5 was the model that made that viable in practice; our own employees were paused before it existed because the output was not worth shipping. Fable 5.1’s positioning is aimed at exactly the failure modes that still show up in that kind of work: shortcut-taking, symptom-patching, going quiet mid-task.

Whether it delivers is a question for the coming weeks, not launch day. Anthropic’s benchmarks are the vendor’s own numbers, and we treat them that way. What we can say today is that the model swapped under the work with nothing lost: every employee’s memory, roles, task board, and org structure carried straight over, because none of that lives in the model.

§ 05 What does not change

Your pricing. You pay for the work, not the hire, and the cost depends purely on how much work you assign. The cache-read cut lowers what it costs us to run long conversations; it does not change what you pay today.

Your setup. There is no model picker and no migration step. If a chat or an employee is on an Agent or Team tier, Core or Max, it is on Fable 5.1.

Your other tiers. Flash and Agent Creative run exactly what they ran before.

Hire an AI employee and it runs on Fable 5.1 from its first shift, carrying what it learns from one shift into the next.

§ 06 What we are watching

Three things over the next few weeks: independent benchmark replication, real cache-read economics on our own bill against Anthropic’s 25% and 45% estimates, and how the Core tiers feel at the lighter reasoning setting, which is the first time Fable has run that way in production for us. Updates land on the tracker page, and anything that changes CellCog’s routing lands here.

Updated September 2, 2026: an earlier version of this post named only the Max tiers. The Core tiers moved the same day; the post now describes the full change.

FREQUENTLY ASKED

Q1 Which CellCog tiers run on Fable 5.1 now?

Four: Agent Core, Agent Max, Team Core, and Team Max. Agent Max moved from Fable 5 with the rest of its operating point unchanged, including its thinking effort, the setting we have run in production since June. Agent Core, Team Core, and Team Max moved from the Claude Opus line. The Core tiers run Fable 5.1 at a lighter reasoning setting and the Max tiers at a deeper one, which is the whole difference between Core and Max now. Flash tiers keep running on Gemini Flash (3.8 Flash since September 2) and Agent Creative keeps running on Claude Opus 5.

Q2 What is Claude Fable 5.1?

Anthropic’s most capable generally available model, released September 1, 2026 alongside a restricted variant, Mythos 5.1, for vetted cyberdefense and life-science organizations. Anthropic positions Fable 5.1 for ambitious, long-running projects and says it avoids easy-seeming shortcuts and fixes root causes rather than symptoms. Its published numbers: 55.8% on Terminal-Bench 4.0 versus 42.0% for Fable 5, and 52.6% on Terminal-Bench-Science 0.1. Our full release record, with the rumor scorecard, is on the Fable 5.1 tracker.

Q3 Why did you switch on day one?

Because Fable 5.1 is a drop-in successor. Anthropic’s model documentation lists the same 1M-token context window, the same 128K output limit, and the same always-on adaptive thinking as Fable 5, so Agent Max’s proven operating point carries over with only the model name changing. For the tiers coming from the Opus line, the same evidence told us the new model was a clean step up at the effort each tier runs. Claude Opus 5 stays in our catalog as the fallback behind it. We did the same thing with Gemini 3.7 Flash on August 13 and with Fable 5 itself in June: when the successor is a clean superset, waiting costs users capability for nothing.

Q4 Does the cheaper cache pricing change what I pay?

Not today. Cache reads are the tokens a model re-reads from a previous turn, and long chats and AI employees re-read a lot of context each turn, so Anthropic’s 75% cut lowers the cost of running that work on our infrastructure. Your pricing is unchanged: you pay for the work, not the hire, and the cost depends purely on how much work you assign.

Q5 How is Fable 5.1 different from Fable 5 in practice?

Anthropic’s positioning is about long-horizon behavior: fewer shortcuts, root-cause fixes, progress updates while it works. Those are exactly the traits that matter for an AI employee holding a task across a shift. Anthropic’s benchmarks are the vendor’s own numbers; independent replication over the next weeks will tell the real story, and we will note what holds up on the tracker page.

Q6 What is the difference between Core and Max now that both run Fable 5.1?

Reasoning depth. Core is the balanced setting for everyday work: the model thinks less per turn, so it is faster and cheaper to run. Max is the deepest reasoning setting for the questions that deserve it. Agent tiers are one agent working in a loop with you or running an AI employee’s shift; Team tiers are several agents challenging each other’s findings, the mode behind our deep research work. All four now run on Fable 5.1.

ALREADY LIVE

Hire the employee, not the model

The model swapped under the work; the memory, roles, and org structure stayed. Hire an AI employee and it runs on Fable 5.1 from its first shift.

KEY POINTS

5 · 5 min full read

• You can now plug any remote MCP server into CellCog on the new MCP servers page: give it a name, paste its https URL, and pick its authentication (none, or an API key).
• CellCog discovers the server’s tools automatically and keeps them in sync daily. Your agents find them in tool search and call them like any catalog tool.
• Every call runs behind your existing Tools approval threshold, with one extra safeguard: calls to your own MCP servers are always treated as at least moderate risk.
• Your MCP servers are private to your account. No other user can see them or call them.
• MCP (Model Context Protocol) is the open standard AI platforms use to connect to tool servers - the same standard behind custom connectors in other major AI products.

§ 01 What MCP is, in one paragraph

MCP - the Model Context Protocol - is the open standard AI platforms use to connect to tool servers. A remote MCP server is an https endpoint that can answer two questions: what tools do you have? and run this tool with these arguments. Because it’s a standard, one server works everywhere the standard is spoken: the custom connectors in other major AI products are MCP servers, growing numbers of vendors ship official MCP endpoints, and teams increasingly wrap internal APIs in one. Now CellCog speaks it too.

§ 02 How it works

1. Add the server. Name, https URL, and authentication: none, or an API key sent as a Bearer token or a custom header. The key is held encrypted by our tools provider - never stored by CellCog, never visible to agents.
2. Tools are discovered automatically. CellCog connects and lists what the server offers - you see the discovered tools before you leave the page. The list re-syncs daily, and a Refresh button re-syncs it on demand.
3. Agents pick them up on their own. Every tool search an agent runs now includes a section listing your MCP servers and their tools. No prompting tricks, no setup in the chat - if the server is connected and Tools is on, your agents know about it.
4. Calls ride the same rails. An MCP tool call looks exactly like any other tool call: the same approval cards, the same per-chat Tools selector (include or exclude servers per chat), the same audit trail.

§ 03 The safety model

Connecting a third-party server is a bigger trust decision than connecting a catalog tool, and the design treats it that way:

• Your existing approval threshold governs every call - nothing about your current setup changes.
• MCP calls are always at least moderate risk. A third-party server writes its own tool descriptions, and those descriptions aren’t vetted by CellCog. So a call to your own MCP server never rides your lowest auto-approve setting: if you review moderate actions, you review these.
• Servers are private to your account. No other user’s agents can discover, inspect, or call them.
• Agents treat server content as data, not instructions. Tool descriptions and results from a third-party server are handled as untrusted data, never as commands. An instruction embedded in a tool result is something the agent reports, not something it follows.

§ 04 Everyone speaks MCP now. The trust model is where products differ.

That is the point of a standard: Claude’s custom connectors, ChatGPT’s custom apps, OpenClaw’s server configs, Grok Bot’s plugins - all MCP underneath. A server you stand up once works across all of them. So the question worth asking is no longer “does this product support MCP?” It is “what happens to trust once a third-party server is inside?”

Grok Bot’s answer, per its own documentation, is account-wide sharing: MCP authentication is shared across Cursor and Grok Bot, all of a user’s Bots share one cloud computer, and files, browser sessions, and command-line credentials on that computer are available across the whole Bot roster. Their docs are commendably direct about the consequence: “Do not use separate Bots as a security boundary.”

We made the opposite bet, because AI employees carry more autonomy than a coding assistant and the blast radius of a bad tool call is bigger:

• A server you add is private to your account, and every call from it is at least moderate risk - it never rides the lowest auto-approve setting.
• You can restrict each AI employee to exactly the tools and credentials it needs: anything you don’t grant never enters that employee’s environment, enforced on the server, not just hidden in the UI.
• Server content stays data. A tool description is never a command.

Neither answer is wrong for its product. A shared computer is convenient when your Bots are one person’s helpers. Per-employee scoping matters when your agents are staff: working different roles, holding different access, reachable by the outside world. If that is your situation, the boundary is the feature. (More on how we think about this: Grok Bot’s security model, explained.)

§ 05 The honest caveats

MCP server support covers no-auth and API-key servers today; OAuth-based MCP servers are on the roadmap. A server’s URL and auth mode are fixed once added - changing either means removing and re-adding it. And the quality of what your agents can do with a server is bounded by the server itself: a well-described tool set gets used well, a vague one less so.

If you’ve been waiting to give your agents a tool the catalog doesn’t have - it’s probably one paste away now.

FREQUENTLY ASKED

6 questions

Q1 What is MCP and why does it matter?
MCP (Model Context Protocol) is the open standard AI platforms use to connect to external tool servers. A remote MCP server is just an https endpoint that lists its tools and executes them on request. Because it’s a standard, one server works across every platform that speaks it - and now CellCog does.

Q2 How do agents use my server's tools?
Automatically. Every tool search an agent runs includes a section listing your MCP servers and their tools. The agent fetches a tool’s schema and calls it exactly like a catalog tool - same approval cards, same audit trail.

Q3 What are the approval rules for MCP tools?
The same threshold you set for Tools governs them, with one addition: a call to your own MCP server is always classified as at least moderate risk, because a third-party server’s tool descriptions aren’t vetted by CellCog. In practice, an MCP call never rides the lowest auto-approve setting: if you review moderate actions, you review these.

Q4 Can other users see my servers?
No. MCP servers are private to the account that registered them. Another user’s agents cannot discover them in search, fetch their schemas, or call them.

Q5 What are the limits?
Up to 10 servers per account and up to 500 tools per server. A server’s URL and auth mode can’t be edited after adding - remove and re-add to change them.

Q6 When should I use an MCP server instead of a key?
Use a key when you have a raw API credential and an agent can script against the service’s API. Use an MCP server when the service (or your team) exposes an MCP endpoint - agents then get proper, schema-validated tools instead of raw scripting.

NEXT STEP

Plug in your first MCP server

If a vendor you use publishes an MCP endpoint, or your team has internal tools worth exposing, connect one and watch your agents pick its tools up in their next search.

Open the MCP servers page

Your AI employees now have browsers of their own.

Cloud Browsers is live for every CellCog account: each AI employee can open a real Chrome that runs on our servers, keeps its logins and tabs between shifts, and works even when your computer is off.

Until now, an agent that needed the web either browsed logged out or borrowed your Chrome through Browse my Chrome, which meant your machine had to be on and your identity did the browsing. That is still exactly right for work on your own accounts. But a standing employee has standing work: the analytics dashboard it checks every morning, the vendor portal it files into, the docs site it reads. Those deserve a browser that belongs to the employee, the same way it already has its own inbox and task board. That is what shipped.

KEY POINTS

7 · 6 min full read

• Cloud Browsers shipped August 31, 2026: every CellCog AI employee can now use a real Chrome of its own, running on CellCog’s servers.
• It is actual Chrome, not a reconstruction: the live view shows the employee’s real browser window, and you can click and type in it yourself.
• Logins and tabs carry across shifts: sign an employee into a tool once and that session is there for future work, even when your computer is off.
• One-click login handoff: when an employee hits a login wall, it sends you a link that opens that exact login page on its browser. You sign in, the employee continues. Credentials never pass through chat.
• Agents drive their own browser with the same commands they already use for Browse my Chrome, so everything an agent learned there applies unchanged.
• Own browser by default, shared by choice: each employee gets its own profile with its own logins, and two employees can even hold different accounts on the same site. Pooling a team onto one shared profile is a deliberate choice you make, not the starting point.
• Browse my Chrome is unchanged: your own Chrome remains the right tool for anything tied to your personal accounts and your machine.

§ 01 A real Chrome, not a remote control

The live view on the Cloud Browsers page streams the employee’s actual browser window: real Chrome UI, real tabs, real address bar. Nothing is reconstructed or approximated. You can take the controls at any moment, click around, open a tab, sign into a site, and hand it back. Whatever state you leave is the state the employee works with.
This matters more than it sounds. When you can see exactly what the agent sees, there is no gap between “what the agent reported” and “what actually happened on the page”. Debugging a stuck flow becomes watching it happen.

§ 02 Log in once, it stays

Logins are the whole point of a persistent browser, so they got the most engineering attention.

When an employee runs into a login wall on a site it expected to be signed into, it does not fail and it does not ask you to paste credentials into chat. It sends you a link. One click opens that exact login page on the employee’s own browser, you sign in the normal way, and the employee continues with the authenticated session. Passwords never transit the conversation.
Once a session exists, it stays with the browser: across shifts, across service restarts, across clean closes and reopens. While the browser runs, its session state is checkpointed continuously, so even a hard crash costs at most the last few minutes of new logins. We tested this the direct way before writing this post: a phone-number login made through the live view, the browser fully closed and reopened underneath it, and the session came back signed in with every tab intact.

§ 03 Built for standing work

Because the browser runs on our servers, it has no dependency on your machine. Your laptop can be closed, your desktop asleep; the employee’s 6am research shift browses anyway. Combined with shift memory, this closes a loop: an employee that learned a workflow on Monday still has the tabs and the logins to run it on Tuesday.

Agents drive their own browser with the same commands they already use for Browse my Chrome. Same navigation, same page reading, same form filling. Everything an agent learned driving your Chrome transfers unchanged to driving its own.

Give an employee the keys to its own tools.

The employees you hire on CellCog come with their own email, task board, memory, and now a browser that holds their working logins. Set an employee up once with the sites its role needs, and every shift after starts already signed in.

§ 04 Two browsers, two jobs

Browse my Chrome is unchanged, and the two are deliberately complementary:

• Browse my Chrome: agents work in your real Chrome, on your machine, with your sessions. The right tool when the task is about your accounts, your open tabs, your local dev server.
• Cloud Browsers: the employee’s own standing browser and identity, always available. The right tool for the employee’s recurring work, research, and role-specific logins.

Profiles can also be shared, deliberately: assign one browser to several employees and a whole team works out of a single set of logins, like a shared workstation that never gets unplugged. Sharers browse one at a time, and the default stays one browser per employee.

§ 05 Separate identities, shared only by choice

The default is the interesting part, so here is the honest contrast. Grok Bot runs every Bot on one shared cloud computer: their own FAQ says Bots “share its files, browser sessions, and logins”, and advises plainly, “Do not use separate Bots as a security boundary” (docs.x.ai FAQ, checked August 31, 2026). Credit where due: that is a candid line to publish, and a shared pool is a coherent design when Bots are interchangeable workers on one person’s tasks.
CellCog starts from the opposite default because employees are not interchangeable. Each one holds a role, and roles come with their own accounts: your support employee signs into the helpdesk as support, your ops employee signs into the same tools as ops. Separate browser profiles mean separate logins, separate tabs, separate cookie stores, and two employees can hold different accounts on the same site. Pooling is available when you want it, per profile, on your call.
We will be equally candid about what this is: practical identity separation, not a hardened security boundary. But practical identity separation is what a business actually runs on. Knowing which employee is signed in as whom, and being able to log one of them out of one site without touching the rest, is the difference between a workforce and a pool.

§ 06 The honest caveats

A browser is an identity. Treat an employee’s browser logins with the same care you give any credential: you decide what it signs into, and approvals still gate anything consequential the employee does on a logged-in site. A hard crash can lose the last few minutes of brand-new logins, though everything checkpointed before it survives. And a browser of their own does not exempt employees from fair play: sites’ rules about automation apply exactly as before, and CellCog employees are built to respect them.

FREQUENTLY ASKED

6 questions

Q1 Do AI employees stay logged in between shifts?
Yes. A Cloud Browser keeps its logins and open tabs across shifts, service restarts, and clean closes. Sessions are checkpointed continuously while the browser runs, so even a hard crash costs at most the last few minutes of new logins.

Q2 What happens when a site asks the employee to log in?
The employee detects the login wall and sends you a one-click link that opens that exact login page on its own browser. You sign in like you would anywhere, and the employee picks the task back up with the authenticated session. No credentials in chat, ever.

Q3 Can I watch and control the employee's browser?
Yes. The live view streams the actual Chrome window. You can click, type, open tabs, and log into sites yourself; the employee sees whatever state you leave behind.

Q4 How is this different from Browse my Chrome?
Browse my Chrome connects agents to your real Chrome on your machine: your sessions, your identity, great for work on your own accounts. Cloud Browsers gives the employee a standing browser of its own that works while your laptop is closed. Same commands on both.

Q5 Can multiple employees share one browser?
Yes, when you choose to. A browser profile can be assigned to several employees, so a team shares one set of logins, like a shared workstation. Each employee sees the same tabs and sessions, and sharers browse one at a time: if one employee is mid-session, the next waits its turn. The default is the opposite: every employee gets a browser of its own.

Q6 Does this let employees get around website rules?
No. CellCog employees follow the same fair-play rules everywhere: no fake engagement, no scraping where it is prohibited, no evading rate limits or bans. A browser of their own does not change what they are allowed to do with it.

Every AI employee on CellCog now has its own computer, and you can open it. A new monitor button in the employee chat header opens a real desktop: wallpaper, a status bar with the employee’s name and shift state, a dock of apps, and windows you can drag, resize, minimize, and stack.

Until now, an employee’s world reached you through panels and tabs: a dashboard here, a files view there, the inbox in its own drawer. All of it worked, and none of it felt like a place. The computer makes it a place. Your employee has a desk now, and looking over its shoulder is a single click.

KEY POINTS

7 · 5 min full read

• Every AI employee now has a visible computer: a desktop with a dock, a status bar, and real windows, opened from a new monitor button in the employee chat header.
• Six platform apps ship with it: Mail, Tasks, Wake, Approvals, Files, and Settings. Each is a real windowed app, not a tab.
• Your employee’s own dashboards appear as apps too, each with its own icon and accent color, and the accent themes the board itself.
• Files now open as independent windows: several files side by side, each closable on its own, instead of one file taking over the panel.
• Mail is your read-only window onto the employee’s inbox: strict chronology under date headers, keyboard triage, and live updates as mail arrives.
• On phones the same computer renders as an app grid with full-screen apps instead of floating windows.
• Nothing about how employees work changed. This is a better way to SEE the work: structured apps over real surfaces, not a pixel stream of a virtual machine.

§ 01 What’s on it

Six platform apps ship with every employee, in the dock from left to right:

• Mail: The employee’s real inbox, read-only: strict chronology under date headers, keyboard triage, live updates as mail arrives
• Tasks: The task board: what’s queued, what’s done, what’s waiting on you
• Wake: Every condition that brings the employee online, built-in and connected, with pause and resume
• Approvals: The standing approvals you’ve granted and what they cover
• Files: The employee’s workspace, browsable, with files opening as windows
• Settings: The employee’s configuration in one place

Each is a real windowed app. Open Mail next to Tasks and read the thread that produced the task. Window positions are remembered per employee, so the desk you arrange is the desk you come back to.

§ 02 Your employee’s own dashboards are apps now

This is the part we find most satisfying. AI employees on CellCog build dashboards for their owners: KPI boards, pipeline trackers, queue monitors. Those dashboards now show up on the employee’s computer as apps, each with its own icon and accent color, sitting in the dock after the platform set. The accent even themes the board itself, so the app you click and the surface that opens feel like one thing.

There’s a quiet inversion in that. Your employee is not just using a computer; it is furnishing one for you. An employee that has been on the job a few weeks greets you with a dock full of things it built.

§ 03 Files open like files should

Opening a file used to take over the entire panel: one file, full width, everything else hidden behind it. Inside the computer, every file opens in its own window. Put the quarterly report next to the spreadsheet it came from and a chart from last month next to both. Each window closes on its own. It sounds small, and it changes how reviewing an employee’s work actually feels.

§ 04 Pixels or product

There are two ways to show someone an agent’s computer. One is to stream the machine itself: Grok Bot, which entered beta this month, shows you a live pixel feed of the shared computer its Bots work on. It is an honest window, and there is something viscerally convincing about watching a screen move.

We went the other way, because everything an employee owns on CellCog is structured: mail is mail, tasks are tasks, files are files. Rendering those as designed apps means you can search the inbox, filter the board, and read the state of a shift at a glance, in both themes, on any screen size. A pixel stream tells you what an agent is doing this second. A computer full of apps tells you where everything stands, which is the question an owner actually walks up with. (Our full comparison of the two architectures is in the Grok Bot security explainer.)

Watch week one instead of imagining it.

The hardest thing to convey about an AI employee has always been what week one feels like. Now you can open its computer and watch the inbox fill and the task board move, from day one. Roles and honest pricing are on the employees page.

§ 05 On your phone

Below tablet width, the same computer renders the way a phone should: an app grid on the home screen, apps opening full-screen with native-feeling navigation, and the dock condensed to essentials. Desktop window layouts are never disturbed by phone visits.

§ 06 The honest caveats

This is a v1, and we’ve shipped it the way we ship everything: real, then richer. Mail is read-only for now (you watch the inbox; you don’t send from it). Some wake-condition editing still lives in the dashboard panel, which coexists with the computer until the computer absorbs it fully. Dock badges for live counts are wired for some apps and coming for others. And the app set will grow: the employee’s memory, activity, and more are natural next tenants.

None of that changes the core of what shipped today: your employee has a computer, it’s furnished with real work, and the door is one click from the conversation you were already having.

FREQUENTLY ASKED

6 questions

Q1 What is the Employee Computer?

A desktop-style home for everything an AI employee owns: its inbox, task board, wake conditions, approvals, files, settings, and the dashboards it builds. It opens inside the employee chat from a monitor button in the header, with real windows you can drag, resize, minimize, and maximize. Window positions are remembered per employee.

Q2 What apps are included?

Six platform apps ship with every employee: Mail, Tasks, Wake, Approvals, Files, and Settings. Beyond those, every dashboard your employee has registered appears as its own app with its own icon, so an employee that built you a KPI board and a pipeline tracker shows them right in the dock.

Q3 Can I open several files at once?

Yes, and this is one of the changes we like most. A file used to take over the whole panel when opened. Now each file opens in its own window, so you can put a report next to the spreadsheet it came from, and close each one independently.

Q4 How is this different from watching an agent's screen?

Some agent products show you the machine by streaming its screen as live pixels. Grok Bot, which entered beta this month, works that way, and it is an honest window. We went the other direction: the employee’s surfaces are structured data, so we render them as designed apps you can search, filter, and read at a glance. Watching pixels tells you what an agent is doing this second; structured apps tell you where everything stands.

Q5 What happens to the employee dashboard panel?

It stays, and both views coexist for now. The computer is the newer, roomier home, and over time it becomes the primary one.

Q6 Does it work on phones?

Yes. Below tablet width the same computer renders as an app grid with full-screen apps, in the style a phone user already knows, instead of shrunken floating windows.

SEE IT

Open your employee's computer

It is the monitor button at the top of any employee chat. No employee yet? Hiring one takes a few minutes, you pay for the work, not the hire, and now you can watch it settle into the role from day one.