Replicate MCP

Replicate hosts a large catalog of machine learning models behind a single HTTP API, and the Replicate MCP server puts that catalog behind the Model Context Protocol so an assistant can browse and run models without a developer writing API calls by hand. The discovery half of the surface covers searching the catalog, browsing curated collections, reading a model's description and version history, and pulling the input and output schema for a specific version, so an agent can work out which parameters a model expects before it invokes anything. The other half is execution. An agent can create a prediction against a chosen model version, poll its status while it runs, retrieve the output, and cancel a run that is no longer needed. Because image and video models return file URLs rather than text, the server also helps surface generated media back into the client, which is what makes conversational prompts like generate an image with this model, then try a different aspect ratio, work end to end. The practical audience is developers and technical teams prototyping with generative models inside an IDE assistant or a chat client. It fits the exploration phase: comparing several image, video, audio, or language models on the same prompt, checking how a version behaves, then lifting the working call into application code. Long running or expensive jobs still belong in your own backend, since prediction state lives with Replicate rather than in the conversation.