Skip to directory

About Replicate MCP

Replicate’s official MCP server for discovering models, retrieving model information, running predictions, and accessing supported Replicate API operations.

What you can use it for

  • Search and compare AI models
  • Read model metadata
  • Create and retrieve model predictions
Connection and setup
Hosting
Local Node.js process launched by MCP client
Transport
stdio
Endpoint
Local process; no hosted endpoint documented
Authentication
Replicate API token supplied as `REPLICATE_API_TOKEN`; tokens are account credentials and must be stored in the client’s secure local environment/configuration.

Requirements

  • Node.js/npm/npx.
  • A Replicate account and API token from https://replicate.com/account/api-tokens.
  • A Codex CLI version that supports stdio MCP servers.

Setup steps

  1. Create or copy an API token at https://replicate.com/account/api-tokens and store it securely as `REPLICATE_API_TOKEN`; never put the token in a catalog, source file, or chat.
  2. Run the official Codex CLI command `codex mcp add replicate --env REPLICATE_API_TOKEN=<your-token> -- npx -y replicate-mcp@latest` using the local secure environment mechanism for the token.
  3. Start a new Codex session and run `/mcp`; confirm tools are listed before use.
  4. Begin with a read-only model search or metadata lookup. Prediction creation can consume account credits, so review cost and ask before running generation.
Tools 3+

Replicate’s Codex setup page shows example tool names `search_models`, `create_predictions`, and `list_hardware` followed by an ellipsis; this is a partial example, not a complete live tools/list.

  • search_models

    Search Replicate models.

    01
  • create_predictions

    Create a model prediction; this may consume account credits.

    02
  • list_hardware

    List hardware options.

    03

This server’s available tools can vary by account, permissions, or configuration. The list shown is not its full runtime catalog.

Limitations and notes
  • Prediction and generation operations can consume paid account credits.
  • The official page describes a remote SSE service and a local npm process; this record follows the explicit Codex setup, which uses the local stdio package.
  • Connection and functionality were not tested in this research pass.
Original sources
  • official vendor codex setup

    Replicate documents the Codex CLI command, local npm package, required token environment variable, and example tools/list output.

  • official vendor mcp reference

    Replicate documents its official MCP server, supported API operations, local package, and remote alternative.

  • official vendor credential guide

    Replicate documents where to create API tokens and secret-handling expectations.

  • official yc company profile

    Y Combinator lists Replicate in its Winter 2020 batch and describes its model-hosting product.

  • official brand asset

    Downloaded the official website’s linked 512x512 PNG favicon; visually reviewed on light and dark halves of the batch contact sheet (company-brand favicon fallback).

  • official tool documentation

    Replicate official MCP reference describes model discovery and prediction operations; hosted tool surface is API-generated and not established as exactly three.

Documentation checked 2026-10-28T00:00:00Z. Connection has not been runtime tested.

Similar tools