OpenWork Release Notes

Work on documents without leaving Workbot. When Documents is enabled for your workspace, preview Drive files before choosing Edit when editing is enabled, keep chat open beside a document, and continue your conversation with the file in view. Google Docs, Sheets, and Slides stay in Google and save to the original; on OpenWork Cloud, Office files can use an editor in your own workspace when enabled.

AI models and providers work more smoothly across OpenWork. Desktop users can use existing Azure Foundry and custom OpenAI-compatible providers on OpenCode v2, see organization model changes in the picker when enabled, and find the same available models in Settings and shortcuts. When AI Gateway access is enabled, cloud agents can use supported models assigned by their organization.

Admins can check setup before work gets stuck. Settings › Diagnostics brings desktop startup, AI provider, connector, and deployment-service checks together. Google Workspace settings can also identify APIs an admin needs to turn on, and Workbot explains when an organization needs an AI provider before it can answer.

Chat keeps files, prompts, and agent work on track. Files saved from chat go to the opened folder, including files created in a conversation’s own folder; links with spaces open correctly, failed writes are labeled as failed, and messages in side chats are easier to start. OpenCode v2 history copies can resume after restart; if a chat is too large to copy whole, its largest outputs or attachments stay in the original history.

Team and Workbot tasks have clearer paths. When enabled, members can connect assigned accounts in Workbot; members with team-management access can open the right team’s settings, and teams can follow tasks started on another computer. When available, admins can set team or organization defaults for new desktop chats and provider-specific spend limits, and inspect Library use and free Auto conversion insights. Workbot’s link-only sharing is gone; authorized invitation emails remain available in Den.

Self-hosting and connected tools get practical fixes. Self-hosters get updated AWS, Google Cloud, and Azure deployment options and clearer service checks. MCP clients can identify OpenWork and receive clearer tool guidance, and the website’s Trust Center links now lead to the hosted Trust Center.

Self-host OpenWork Web and team tools on AWS. For self-hosters: run OpenWork Web with sandboxed browser workspaces, and deploy Slack, Workbot, and Automations that run without a desktop. AWS deployments can use existing networks and load balancers, and can connect to an existing cluster or runner fleet.

Connect OpenCode 2 to your organization. If you use OpenWork from OpenCode: the opencode-openwork plugin brings in your organization’s AI Gateway models, tools, and skills after sign-in, without the desktop app. You can also upgrade desktop chats to OpenCode 2 with a backup first; its conversations can use the built-in browser and discover connected tools reliably.

Admin controls and reporting

Admins get clearer controls and reporting. Organization Analytics is back, with restored usage and workflow history where available. When enabled, admins can also report on skills, plugins, and connectors, set member, admin, and team permissions, and choose a default model for cloud agents; Audit Logs now cover more organization activity.

Chat and daily work

Chat and daily work recover more smoothly. Pasted text and images stay at the cursor, reopened Apps load their data, migrated chats with a missing agent can answer again, and OpenWork keeps the Working status visible during quiet stretches. Ollama keeps multiple selected models and their thinking choices; Drive uploads respect your chosen Google account.

Calendar, Workbot, and connections

Calendar, Workbot, and connections do more. When Calendar features are enabled, add or edit Automations alongside meetings; Workbot can recover stopped tasks. When enabled, admins can check when connections were last verified, and compatible MCP clients can subscribe to remote-session updates.

Familiar pages and searches

Some familiar pages and searches have changed. Organization dashboards and admin-assigned dashboards in Desktop are gone, while personal dashboards remain. “Search sessions” now searches session titles and workspace names, not message text; for organizations with OpenWork Models, Auto is replaced in the model picker by GPT-6 Luna.

Move your Cowork setup with one prompt

Ask the agent you already use to follow openworklabs.com/migrate.md to bring your Claude Cowork and Claude Code plugins and skills into OpenWork. Re-running picks up changes without duplicates, reuses connections your organization already has, and explains when a connection still needs setup.

Give each teammate their own connector key

For admins: external connections can use individual keys instead of one shared key. Members can add or replace their key from the Library or Den, using a masked entry form; desktop chat can guide them to add one when needed. For admins who use LiteLLM, connecting your own proxy is also available as a preview on request.

Auto and Apps are available to more organizations

Free Auto no longer needs an organization-by-organization rollout, and every organization can build Apps without a platform admin enabling them. Auto can also read images you attach or that tools return, and signed-out Linux desktops no longer need a system keyring to use it.

Self-hosting gains a Google Cloud Run starting point

For self-hosters: starter deployment files and instructions cover the Den API, Den web app, and migration job on Cloud Run. Old self-hosted Den settings that had no effect are no longer read.

Everyday work is clearer and more dependable

Your first organization message now gets an answer before you have a workspace; videos made in a conversation’s work folder play in chat; and stopping a run no longer leaves a bold “Step interrupted” row. Agent-led installs open on Setup complete without requiring Node first, and teammate invites survive workspace claiming. Model names and provider details are clearer. Workbot uses one quiet activity line and can react with emoji, while Slack runs wait for slow desktop work instead of reporting that they are stuck.

Some desktop options have changed or are gone

Workspace sharing and remote-workspace connections have been removed. Organization Analytics is gone, while usage reporting remains in AI Gateway. Plugins now come from your organization’s Library instead of being installed into workspaces, and the “From GitHub” plugin import is gone. The Library no longer has a card view. Automation choices now say Desktop or Cloud, and computer use has been removed from that editor.

Find shared connectors that still need your sign-in

Shared connectors are easier to find. The Library now has a “Needs sign-in” filter for connectors shared with you that still need your sign-in. In Den, a sidebar dot also points out when something is waiting on you.

Self-hosted first runs are faster

For self-hosters: openwork-server web opens your first folder in about a second on a fresh install, instead of taking 10+ seconds on Windows. It also no longer fails with “resource busy or locked” while installing the engine.

Side chats no longer crowd the sidebar

Their chips have been removed from session rows, leaving more room for session names. Open a session to see its side chat.

Workflows and artifacts can use your connected tools

Connected tools work in Workflows and artifacts. Live artifacts, on-demand Workflows, and scheduled Automations can now use connection tools and perform write actions. Large tool inputs, such as Notion database queries, also work.

Work sent to your desktop reports back

If you use OpenWork from Slack or another connected app, you can now follow a desktop task’s progress, read its conversation, send a follow-up, or stop it. Slack reports the result in the original thread, and tasks appear as normal chats using your chosen model.

Auto works in more situations

Eligible organization members can use Auto alongside other models, and signed-out Auto is available to OpenWork builds and compatible clients. Deployments can disable Auto while keeping other models available.

Chat and the built-in browser are clearer and more dependable

Model choices explain why a model is unavailable; OpenCode v2 now shows a Connect card when a service needs sign-in. The browser explains when a page cannot connect and offers Reload, and HTML file previews are isolated from the rest of OpenWork.

For admins and self-hosters

Admins can choose “Only models you provide,” enable Slack progress updates, and review free Auto usage. Self-hosters can deploy with draft Terraform modules for AWS ECS Fargate or Kubernetes, including an existing ECS cluster.

Other changes

OpenWork-built items are now called artifacts. Browser login sync and its setup controls are gone; sign in directly in the built-in browser. Organization-managed Dashboards are disabled by default and must be enabled per organization.

Enterprise sign-in and version pinning are back, plus Amazon Bedrock and model shortcuts

Organization sign-in and allowed versions work again. OpenWork Enterprise and OpenWork Cloud desktops once again ask people to sign in to their organization before they start working, and desktop updates skip any version an admin removed from Allowed desktop versions. Other desktop policies stay saved but are not enforced on the desktop while they are redesigned.

Amazon Bedrock in AI Gateway. For admins: add Amazon Bedrock and Amazon Bedrock (OpenAI) with an AWS region and access keys, and reuse the saved keys for the second one. Anthropic effort choices are kept through the gateway.

Switch models with a key. In Settings › Keyboard shortcuts you can bind a key to a saved model with its reasoning level and Fast. ⌃⇧F (Ctrl+Alt+F on Windows and Linux) turns Fast on or off, and the model name shows ”· Fast” while it’s on.

Chat keeps its place. Editing a message while a reply is running replaces it instead of queueing the edit. Conversations stay in their folder when an agent changes directory, and questions and approvals no longer disappear. You can switch between OpenCode engines from the command palette and choose whether to copy older chats across. Agent tool work is summarized by what it did instead of raw tool names.

Simpler connector setup. For admins in OpenWork Cloud: API-key, custom sign-in, Google Workspace and Microsoft 365 connectors use the new setup flow and can be edited in place, and Plugins has clearer filters. Apps like Claude Code, Codex and Cursor can connect to OpenWork Connect without registering first, failed sign-ins explain what to do next, and the MCP docs have one-click Cursor and VS Code install links.

Also in this release. Windows installers are published again. Updating no longer fails with a checksum error after a release is re-run. Update controls in Settings are more compact, a second OpenWork profile no longer stops the first one’s engine, and the website shows OpenWork’s completed SOC 2 Type II audit.

Your Library works fully on the desktop.

You can now add, share, edit and delete skills, connectors and plugins without leaving the app. A new Connectors tab lists your connectors, adding one starts from a catalog with service logos, and an MCP server’s address is checked before it’s saved. Each item’s page has Chat, which starts a task with that item ready to use, and Share on items you own. When a skill shapes an answer, chat names the skill. Signed out, the Library shows one clear Sign in to OpenWork Cloud button instead of an “Add to library” button that did nothing. Lists also load faster.

One + menu in the composer finds everything.

The paperclip and tools buttons are now a single + menu that opens on a search across your files, skills, connectors, plugins and commands. It still finds what you meant when you make a typo (“hubsopt” finds HubSpot). Attaching a file is the first option, or press ⌘U (Ctrl+U on Windows and Linux).

Usage limits are clearer.

When you reach your organization’s usage limit, the chat says which limit you hit and when it resets, and you can ask for more right there. The account menu shows how much of each limit you have left, and the new Settings › Usage page has the details. The chat no longer also shows a misleading “too many requests” error under the limit.

AI Gateway is rebuilt and available to every organization.

For admins: AI Gateway is now a regular page in the sidebar. Add a provider from a catalog, then set its key, who can use it and which models in one form. Overview shows who spends what and on which models, with a page for each person. Limits lets you cap spend per day, week or month for everyone, a team or one person, and you can undo a deleted limit. Users & Teams shows everyone’s providers, limits and last 31 days of spend. Mistral is now supported. Members can sign in to Gemini and Claude on Google Cloud with their own Google account from the new Models section of My Library, which replaces the old My Model Connections page.

Connectors and plugins are simpler to manage.

For admins in OpenWork Cloud: add a connector in a few steps, choose whether each person signs in or one account is shared by everyone, and pick which teams get it. Add any MCP adds a server that isn’t in the catalog, or you can paste its address into the search. Manage › Plugins lists every plugin with who has it. You can create one and give it to teams right there, then search by name or filter by team or person. Removing a connector or plugin now asks first, and switching a connector to one shared account warns you before it signs you out. Owners can also approve their self-hosted OpenWork web address in Org settings so members can sign in there, even if they belong to several organizations.

Also in this release.

The first time you click a link, OpenWork asks whether to open it in OpenWork or your usual browser and can remember your choice. You can change it later in Settings › Preferences › Browser, or right-click a link to choose once. Session search and notifications sit next to the OpenWork name at the top of the sidebar, and window controls look native on macOS, Windows and Linux. The built-in browser no longer has a Suspend button. After you sign in, Settings › AI providers works right away, even before you create your first workspace. If you use OpenWork from Claude Code, Cursor or Codex, agents can update shared skills. When a long document or email is cut short, they’re told the full length and how to read the rest.

Self-host OpenWork with one command.

For self-hosters: npm i -g openwork-server && openwork-server web now runs the full OpenWork web app on a home server or VPS. It downloads the matching engine on first run and keeps its access tokens across restarts, and npm i -g openwork-server@latest updates it. Before, you had to clone and build the repo, install the engine yourself, and recreate tokens after every restart.

Connecting an account in chat no longer gets stuck.

When the agent needs you to connect a service, you now always get a clear card to connect or skip, with Try again if sign-in doesn’t finish. Before, chat could wait forever on “Checking connection request…”.

Chat replies are simpler.

Creating or sharing a skill, asking for a mockup, being asked for an environment variable, and browsing connectors now reply in plain text instead of embedded cards. The connection card is the only one that stays.

Your email stays out of the sidebar.

The account row now shows your name and “OpenWork Cloud” instead of your email address. Your email is still at the top of the account menu.

If you use OpenWork from Claude Code, Cursor, or Codex: agents can now list your organization’s skills and open any of them directly.

Before, they had to guess search keywords, so organization skills often went unused.

Clearer pricing everywhere.

The website, docs, and sign-up now all show the same plans: Free, Team at 10𝑝𝑒𝑟𝑠𝑒𝑎𝑡 𝑝𝑒𝑟 𝑚𝑜𝑛𝑡ℎ, and Enterprise at 40 per user billed annually, with Cloud Computer and OpenWork models as add-ons. The website and docs also have refreshed branding.

Agent sessions recover more reliably and provider connections stay consistent

Conversations now reopen and continue more reliably, with history pages filling correctly and background streaming preserved.

Mobile chat is easier to use, with steadier submission, scrolling, keyboard, composer, and sidebar behavior.

Provider connections and cloud workspace startup now handle credentials, readiness, and recovery more reliably.

Removed custom agents and added clearer linked task references and alpha build information.

Desktop connection status now uses one clear native card, while sign-in and managed credentials recover more reliably.

Browser tasks can use site tools, page controls, tab suspension and reload, selected logins, and user takeover.

Agent runs recover more smoothly from stalled sends, interruptions, restarts, and follow-up turns, with drafts and sent messages staying visible.

Computer Use adds hosted approval, native previews, scoped sessions, and continuation after takeover or refresh.

Organization access, workflow apps, analytics, onboarding, and desktop policies are clearer and more dependable.

The Working indicator now tells the truth: it stops when a run errors or the engine is unreachable, and shows minutes once a run passes the first hour.

Sessions reconcile more cleanly — completions are detected while you’re working elsewhere, queued messages send even when the session is out of view, and archiving works from any workspace.

Cloud providers now connect using the credentials saved on your desktop, and Google Workspace retrieval is more resilient.

Published the SOC 2 badge and compliance status on the landing page.

Hardened security by fixing static-analysis findings and isolating provider sync contexts, alongside internal test and release tooling improvements.

OpenWork can now import and run portable Agent Plugins, bringing Anthropic-compatible plugin packs into the desktop app.

Automations can be pinned to a specific workspace, and remote-session commands now arrive in native local sessions on your desktop.

Fixed request storms and reload loops across Settings, the Library, the composer, and desktop policies, so long sessions stay fast.

Gmail replies support collapsible quoting and attachments, and OpenWork Cloud members get complimentary OpenWork Web access.

Strengthened sign-in and access security and made the desktop recover cleanly from socket and renderer crashes.

Session files now live in the artifact rail instead of interrupting the chat transcript, with safer badge layout and responsive behavior on narrow screens.

Markdown can render Mermaid diagrams safely, and code blocks offer a word-wrap control for long lines.

Agent runs recover more reliably from idle admissions, queued follow-ups, event-stream restarts, Settings visits, and large workspace session histories.

OpenWork Cloud adds per-member managed-model credentials, paid browser access, desktop delivery for remote-session commands, and more reliable worker recovery.

Electron file transfers now work correctly across the remote bridge, and LiteLLM examples include synchronized model metadata.