Replicate MCP Server

Replicate hosts thousands of open-source and commercial models behind one HTTP API — Flux and SDXL for images, video models, speech and transcription, embeddings, and language models — billed by the second of compute. This MCP server puts that whole API in front of an agent, so a model can find the right model, read its input schema, and run it without you writing any glue code.

What it exposes

The server is generated from Replicate's HTTP API rather than hand-written around a fixed tool list, which is the interesting design decision here: it mirrors every operation the API supports and picks up new ones automatically as Replicate ships them. In practice that means model discovery (models.search, models.list), schema inspection (models.get), and the prediction lifecycle (predictions.create, predictions.get) — plus collections, trainings, deployments and account operations from the same surface.

Because the tools carry each model's own input schema, an agent asked for "a 16:9 photorealistic image of X" can look up a suitable model, read what parameters it accepts, and submit a valid prediction, instead of guessing at field names.

Running it

There are two ways in. The remote server at `https://mcp.replicate.com` is the simplest — nothing to install, and clients that speak remote MCP can connect directly. The local server ships as the replicate-mcp npm package and runs under npx, needing only Node.js and a REPLICATE_API_TOKEN from your Replicate account dashboard. An experimental Code Mode additionally requires Deno and runs sandboxed TypeScript, which is worth it when a task means chaining several predictions rather than firing one.

Worth knowing

Predictions cost real money — Replicate bills per second of GPU time — so an agent with this server wired up can spend from your account. Scope the token accordingly, and be deliberate about which models you let it reach. The package is Apache-2.0 licensed and published by Replicate.