Replicate Release Notes

Replicate now publishes agent skills, a collection of markdown instruction files that give coding assistants expert knowledge about working with AI models on Replicate.

Skills cover model discovery, comparison, and execution via the API, along with detailed prompting techniques for image generation and video generation models. They follow the open Agent Skills spec and work with Claude Code, OpenCode, OpenAI Codex, and other compatible tools.

Install

npx skills add replicate/skills

This installs all of Replicate’s skills into your project and configures them for your coding assistant automatically.

Skills and MCP

Skills are complementary to Replicate’s MCP server. MCP gives your coding assistant API tools. Skills give it knowledge about how to use those tools well: which models to choose, how to write prompts, and what tradeoffs to consider.

For more details, see the agent skills reference or the GitHub repository.

How it works

Set allow_fallback_model to true when calling the API. If Nano Banana Pro hits a rate limit, it tries to generate the image with Seedream 5.0 lite instead. For certain inputs, for example if the aspect ratio isn’t supported, the original rate limit error is returned.

The fallback is off by default. If you don’t set allow_fallback_model, nothing changes — you’ll get a rate limit error when Google’s API is at capacity.

When the fallback is triggered, your logs still show a prediction to Nano Banana Pro. You can tell the fallback was used by checking the resolution field in your output — it says "fallback" instead of the actual resolution. You’re charged the cost of the fallback model, not Nano Banana Pro.

Limitations

Our current fallback model, Seedream 5.0 lite, doesn’t support all the same options as Nano Banana Pro:

• Seedream 5.0 lite doesn’t support 1K resolution. If you request 1K, the fallback generates at 2K and downscales the result.
• Seedream 5.0 lite doesn’t support 4K resolution. If you request 4K, the fallback won’t be used and the original rate limit error is returned.
• Seedream 5.0 lite doesn’t support the 4:5 and 5:4 aspect ratios. Requests with these ratios won’t fall back and will return the original rate limit error.

MCP server discovery

Replicate’s MCP server can now be discovered automatically through the official MCP Registry.
We added a /.well-known/mcp/server.json endpoint that publishes metadata about the MCP server. This follows the server.json specification from the Model Context Protocol.

How discovery works

The MCP Registry is the official metadata repository for MCP servers, backed by Anthropic, GitHub, and Microsoft. It doesn’t host code—just metadata that describes where to find servers and how to install them.
When you publish a server.json file at /.well-known/mcp/server.json, the Registry can discover your server automatically. MCP clients then use the Registry to find and install servers.

Clients with built-in discovery

A few MCP clients have built-in marketplaces or directories:

• VS Code has the best Registry integration. Enable chat.mcp.gallery.enabled in your settings, then search @mcp in the Extensions view to browse and install MCP servers.
• Claude Desktop has a curated extensions directory at Settings > Extensions > Browse extensions.
  Other clients like ChatGPT, Cursor, and LM Studio require manual configuration—you add the server URL or edit a config file yourself.

Code mode option

The metadata also exposes the --tools flag, which lets you choose between standard tools (all) or code mode (code) when installing.

Filter list predictions API by source

You can now filter the list predictions API endpoint to show only predictions created through the web interface.

Use the source query parameter with a value of web :

curl -s \
-H "Authorization: Bearer $REPLICATE_API_TOKEN" \
"https://api.replicate.com/v1/predictions?source=web"

This is useful if you want to see predictions you created using the playground or other parts of the Replicate website, separate from predictions created programmatically via the API.
Note: When filtering by source=web , results are limited to predictions from the last 14 days.

Web

• Added created and last updated dates on model pages
• Improved pricing display on model pages with clearer messaging for bring-your-own-token models and better visual consistency for single and multiple pricing tiers
• Made FLUX.2 the default model in the playground
• Added FLUX.2 models (pro, flex, and dev) to the playground
• Improved homepage hero section with clickable hero images that link to model pages, better mobile responsiveness, and reduced bundle size
• Improved mobile homepage hero legibility by adjusting gradient overlays
• Fixed model card text truncation and spacing on the explore page at smaller viewport sizes
• Updated playground to better adhere to platform rate limits
• Improved data retention banner messaging to clarify that failed predictions are not retained indefinitely
• Fixed support form reliability issues
• Fixed an issue where the navigation may have rendered incorrectly on mobile devices

Docs

• Published blog post on how to run FLUX.2 on Replicate, Black Forest Lab’s most advanced image generation model
• Published blog post on how to run Isaac 0.1 on Replicate, an open-weight vision-language model

Web

• Added approximate cost display to predictions and trainings on dashboard and predictions pages, showing how much each run costs
• Launched an updated homepage hero section and navigation
• Improved display of code snippets on the Replicate homepage
• Made Nano Banana Pro the default model in the playground
• Improved site search sorting, and improved keyboard shortcut handling

Cog

• Released Cog v0.16.9 with a fix for x-order bug

Docs

• Published blog post on how to prompt Nano Banana Pro with guidance on using its logic, text rendering, character consistency, and world knowledge capabilities

Code mode

Replicate’s local MCP server now supports an experimental “code mode” that allows language models to write and execute TypeScript code directly in a sandboxed environment.

Instead of exposing individual API operations as separate tools, Code mode provides two tools: one for searching SDK documentation, and another for executing TypeScript code using the Replicate SDK within a Deno sandbox. The model uses a built-in docs search tool to learn how to write code against the SDK. This approach is more efficient for complex workflows that involve multiple API calls, as it reduces context window usage and allows the model to write custom logic that calls multiple methods and returns only the final results.

To use code mode, start the local MCP server with the --code-mode flag:

npx -y replicate-mcp@alpha --tools=code

Here’s how to add the local Replicate MCP server in code mode as a tool in Claude Code:

claude mcp add "replicate-code-mode" --scope user --transport stdio -- npx -y replicate-mcp@alpha --tools=code

Code mode is currently experimental and subject to change. It requires Node.js and Deno to be installed locally. Remote cloud sandboxing support is planned but not yet available.

To get started with Code mode, see the Code mode documentation or visit the demo GitHub repo.

Web

• Added video support to the before/after slider, allowing side-by-side comparison of video outputs from models like video upscaling and style transfer
• Added pixelated image rendering with size controls (1x, 2x, fit) in the playground for pixel art models
• Improved support for aspect ratios in prediction outputs
• Overhauled focus states across the Replicate UI library
• Add search to Deployments
• Fixed a bug that caused model input fields to display in a different order
• Added FAQs to every collection page

Platform

• Launched deployment setup monitoring for enterprise customers with automatic email notifications when deployments fail setup and customizable setup timeouts

Docs

• Published blog post on how to run FLUX.2 on Replicate, Black Forest Lab’s most advanced image generation model
• Published blog post on how to run Isaac 0.1 on Replicate, an open-weight vision-language model

Enterprise deployment controls

Setup failure notifications

If a deployment fails during the model’s setup function, we’ll notify you via email. This helps you catch issues with your model earlier, before your users reach out.

If you use Slack, you can configure these emails to be sent to a Slack channel using their send emails to Slack feature.

Custom setup timeouts

You can customize the timeout for your deployment’s model setup function. The default timeout is 10 minutes, which works for most models. But if your model needs to download large files, load trained weights, or perform other expensive initialization operations, you can give it more time before we mark it as failed.

You can configure both of these in your deployment settings.

Read more in our deployment monitoring docs or reach out to [email protected] to learn about our enterprise plans.

SDK

• Our Python SDK public beta is out now, making it even easier to run AI models from your Python code with full support for every operation in our HTTP API

Web

• Improved playground handling of different output and streaming types
• Fixed search keyboard shortcut in docs to use Cmd+/ on Mac and Ctrl+/ on Windows for consistency across platforms

API

• Launched prediction deadlines so you can automatically cancel predictions that don’t complete within a specified duration
• Added source field to prediction API responses to indicate whether predictions were created via web or api

Docs

• Split deployment monitoring into its own docs section with detailed information about metrics, GPU memory monitoring, and performance tracking
• Published blog post about how to prompt Veo 3.1 with guidance on reference images, frame control, and image-to-video features

How it works

You can now set a deadline to automatically cancel a prediction if it doesn’t complete within a specified duration. This is useful when you’re building real-time or interactive experiences, like a virtual try-on experience for an online clothing store. In this case, shoppers have usually moved on if an image takes more than 15 seconds to generate.

Set a deadline by including a Cancel-After header when creating a prediction. See our docs for details on the header format.

Here’s an example that sets a 1 minute deadline:

curl -X POST \
-H "Authorization: Bearer $REPLICATE_API_TOKEN" \
-H "Cancel-After: 1m" \
-H "Prefer: wait" \
-H "Content-Type: application/json" \
-d $'{
  "input": {
    "prompt": "The sun rises slowly between tall buildings. [Ground-level follow shot] Bicycle tires roll over a dew-covered street at dawn. The cyclist passes through dappled light under a bridge as the entire city gradually wakes up."
  }
}' https://api.replicate.com/v1/models/bytedance/seedance-1-pro/predictions

What happens when a deadline is reached

Replicate sets the prediction’s status to aborted if the deadline is reached before it starts running, and canceled if the deadline is reached while it’s running.

For public models, you’re only charged for predictions with a canceled status, not for aborted ones.

Deadline vs sync mode wait duration

Prediction deadlines and sync mode serve different purposes. Use prediction deadline (Cancel-After header) to control when the prediction itself should be canceled. Use sync mode (Prefer: wait header) to control how long the HTTP request stays open waiting for results.

You can also use both together. In the previous cURL example, Prefer: wait defaults to 1 min and we’ve explicitly set Cancel-After to 1 min. This means that the HTTP request will stay open for 1 minute to wait for results, after which the prediction will be canceled, even if it has not completed.

Alternatively, setting Cancel-After: 1m and Prefer: wait=10 means that the request returns after 10 seconds. If the prediction is still running, you’ll get an incomplete prediction object, and the prediction will continue to run until it completes or is canceled at the 1-minute deadline.

Read more in the docs:

• Create a prediction: Prediction deadlines
• Prediction lifecycle: Timeouts

Web

• Improved speed and performance across Replicate
• Updated dashboard UI with new navigation components for improved consistency
• Improved collection API filtering for better performance
• Made the Official label on model pages clickable, linking to the official models documentation

API

• Launched the ability to update model metadata via API with PATCH requests to update descriptions, README content, and links
• Added sorting options to the models.list API to sort by model creation date or latest version date

Docs

• Expanded docs sidebar by default to make navigation easier
• Added comprehensive documentation about rate limiting when you have no payment method
• Published a blog post about IBM Granite 4.0 models now available on Replicate
• Updated getting started guides

Platform

• Added ability to download invoices from billing settings for both monthly billing and credit purchases

Cog

• Released Cog v0.16.8 with registry migration support and credential fallback
• Updated FastAPI requirement to support versions up to 0.119.0
• Fixed Go build issues with version control information injection
• Upgraded test dependencies including TensorFlow updates

API changes

The GET /v1/models endpoint now supports sorting with the sort_by and sort_direction query parameters. This makes it easier to fetch the newest models via the API.

Sorting options

• model_created_at: Sort by when the model was first created
• latest_version_created_at: Sort by when the model’s latest version was created (default)

Sort direction can be asc (ascending, oldest first) or desc (descending, newest first).

The default behavior remains unchanged: models are sorted by latest_version_created_at in descending order (newest versions first).

Update model properties via PATCH

You can now update model properties using the API with a PATCH request to /v1/models/{owner}/{name}.

You can update the following properties:

• description - Model description
• readme - Model README content
• github_url - GitHub repository URL
• paper_url - Research paper URL
• weights_url - Model weights URL
• license_url - License URL

Example cURL request:

curl -X PATCH \
  https://api.replicate.com/v1/models/your-username/your-model-name \
  -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "description": "Detect hot dogs in images",
    "readme": "# Hot Dog Detector\n\n🌭 Ketchup, mustard, and onions...",
    "github_url": "https://github.com/alice/hot-dog-detector",
    "paper_url": "https://arxiv.org/abs/2504.17639",
    "weights_url": "https://huggingface.co/alice/hot-dog-detector",
    "license_url": "https://choosealicense.com/licenses/mit/"
  }'

See the API reference for full details.