The useful pairing is search_models with create_and_poll_prediction: describe the kind of model you want, get one, run it, and see the image — without opening Replicate to copy a version string. The templates matter more than they look, since they save the assistant from inventing plausible-sounding image parameters. Worth knowing before you build on it: the author has stopped maintaining this and Replicate now ships an official server.
A community MCP server for Replicate, covering model discovery, prediction runs and viewing the images that come out. The README marks it as no longer in active development and points to Replicate's own official MCP server; the code remains installable and is a reasonable base to fork.
- Find something to run — search_models by meaning, list_models by owner, and get_model for versions and details
- Browse Replicate's curated groupings — list_collections and get_collection
- Run a model — create_prediction returns immediately, create_and_poll_prediction waits until it finishes
- Follow work in flight — get_prediction for status, list_predictions for what you ran recently, cancel_prediction to stop one
- See the result — view_image opens generated output in your browser, with clear_image_cache and get_image_cache_stats managing what is kept
- Presets so a prompt does not have to carry every parameter: quality levels from draft to extreme, styles such as photographic, cinematic or anime, and sizes including square at 1024x1024 and portrait at 832x1216
- create_prediction accepts a webhook_url, so a long run can call back instead of being polled
A Replicate API token, set as REPLICATE_API_TOKEN in the client config or the environment. Install globally with npm install -g mcp-replicate, or run it with npx. Replicate bills for the compute a prediction uses, and its rate limits apply to this the same as to any other client.
One command plus a key — npm install -g mcp-replicate, then supply credentials