Replicate MCP server: remote and local setup for Claude

Yes: Replicate has an MCP server. You can add its hosted version at https://mcp.replicate.com/sse to Claude, Cursor, or another MCP client, or run the replicate-mcp npm package locally with npx. Either way it authenticates with a Replicate API token, and it can search, compare, and run models through Replicate's HTTP API.

The Replicate facts come from its MCP docs page, the setup page at mcp.replicate.com, and its billing docs, read on 2026-09-28. Sume has no connection to Replicate's server; Sume vs Replicate compares the platforms.

Should I use the remote or the local Replicate MCP server?

Replicate recommends the remote server for most users. You add the hosted URL to your app, and a web-based authentication flow asks for a Replicate API key for the server to use on your behalf; the hosted server is updated automatically. The local server runs the replicate-mcp npm package on your machine, configured in each app with a REPLICATE_API_TOKEN environment variable. Replicate notes that Claude Desktop runs local servers only in the desktop app; the claude.ai web app needs the remote server. Local vs remote MCP server explains the trade-off in general.

How do I add Replicate MCP to Claude Code, Cursor, or Codex?

Create a Replicate API token first. Then follow the entry for your client:

Why is the Replicate MCP URL an SSE endpoint?

The hosted URL Replicate gives ends in /sse, and its Claude Code command passes --transport sse. Claude Code's docs call SSE deprecated and say to use HTTP servers where available. They also say claude mcp add --transport http tries HTTP first and switches to SSE when the server doesn't accept it, from Claude Code v2.1.265. MCP SSE vs Streamable HTTP covers the two transports.

What can it do, and how is it billed?

Replicate says the server supports all of the operations in its HTTP API: searching models (models.search), comparing them (models.list), fetching metadata (models.get), running models (predictions.create), and fetching predictions (predictions.get). An experimental Code mode instead gives the model two tools, a docs search and a TypeScript runner in a Deno sandbox, and supports local execution only for now.

Running a model is a prediction billed to the account whose token the server uses. Replicate's billing docs describe pay-as-you-go billing: on a public model you pay only for the time it is active processing your requests. The page says failed runs aren't charged, with exceptions: private models and deployments bill failed and canceled runs for active instance time, and a failed model that calls other models is billed for its run time plus the downstream models it called. Canceling a run on an official model may still be charged.

How does Sume's hosted MCP server handle auth and spend?

Differently. Sume's server at https://mcp.sume.com/mcp takes an OAuth sign-in or a Sume API key (OAuth and API keys). Under OAuth the session is read-only by default: mutating and paid tools stay hidden until you turn Write on at consent, and an API key sees the full tool set. Spend comes from the Sume workspace wallet, job by job; Connect Claude Code, Cursor, or Codex to Sume covers setup and per-call spend controls (Tools and gates). Sume basics says hosted MCP still works but is not the primary path today.