Replicate Anywhere

An MCP (Model Context Protocol) server that enables AI assistants to search, discover, and run any model on Replicate. No hardcoded model lists - just describe what you want and let the AI find and run the right model.

Features

• 🔍 Smart Model Search - Find models by fuzzy name matching (e.g., "flux", "stable diffusion", "nano banana pro")

• 🤖 Automatic Model Discovery - AI searches first, asks questions later

• 📋 Parameter Detection - Automatically retrieves and understands model input schemas

• 🖼️ Inline Image Display - Image outputs are formatted as markdown for inline display

• ⏱️ Async Prediction Handling - Long-running predictions return status URLs instead of timing out

• ✅ Prediction Status Checking - Check on running predictions that haven't completed yet

Related MCP server: OpenRouter MCP Server

Installation

Prerequisites

• Node.js 18+

• A Replicate API token

NPM (Global)

npm install -g replicate-anywhere

From Source

git clone https://github.com/fifthseason-ai/replicate-anywhere.git
cd replicate-anywhere
npm install
npm run build

Configuration

Environment Variables

MCP Client Configuration

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "replicate-anywhere": {
      "command": "npx",
      "args": ["-y", "replicate-anywhere"],
      "env": {
        "REPLICATE_API_TOKEN": "r8_your_token_here"
      }
    }
  }
}

LibreChat

Add to your librechat.yaml:

mcpServers:
  replicate-anywhere:
    type: stdio
    command: npx
    args:
      - -y
      - replicate-anywhere
    env:
      REPLICATE_API_TOKEN: "${REPLICATE_API_TOKEN}"

Docker

services:
  replicate-anywhere:
    build:
      context: ./replicate-anywhere
    environment:
      REPLICATE_API_TOKEN: ${REPLICATE_API_TOKEN}

Tools

search-models

Search for AI models on Replicate by name or description. This tool is designed to be called first when a user mentions any model name.

{
  "query": "flux pro"
}

get-model-info

Get detailed information about a specific model, including its input parameters schema.

{
  "owner": "black-forest-labs",
  "name": "flux-pro"
}

run-model

Run a prediction on any Replicate model.

{
  "model": "black-forest-labs/flux-pro",
  "input": {
    "prompt": "A beautiful sunset over mountains",
    "aspect_ratio": "16:9"
  }
}

list-models

List public models on Replicate (paginated).

{
  "cursor": "optional_pagination_cursor"
}

check-prediction

Check the status of a running prediction.

{
  "prediction_id": "abc123xyz"
}

Usage Examples

Generate an Image

User: "Generate an image of a cat wearing a space helmet using flux"

The AI will:

1. Call search-models with query "flux"

2. Call get-model-info to get parameters for the best match

3. Call run-model with appropriate parameters

4. Return the image inline (markdown formatted)

Use a Specific Model

User: "Use stable diffusion xl to create a cyberpunk cityscape"

The AI will search for "stable diffusion xl", find stability-ai/sdxl, and run it.

Check a Long-Running Prediction

User: "Check on my prediction abc123"

The AI will call check-prediction to get the current status and output if complete.

Output Formatting

Images

When a model returns image URLs, the output is automatically formatted as markdown:

**Generated Image:**

![Generated Image](https://replicate.delivery/...)

**Direct link:** https://replicate.delivery/...

Other Outputs

Non-image outputs are returned as JSON.

Development

# Install dependencies
npm install

# Build
npm run build

# Watch mode
npm run dev

# Run locally
REPLICATE_API_TOKEN=your_token npm start

Architecture

┌─────────────────┐     ┌──────────────────┐     ┌─────────────────┐
│   AI Assistant  │────▶│ replicate-anywhere│────▶│  Replicate API  │
│  (Claude, etc.) │◀────│    MCP Server    │◀────│                 │
└─────────────────┘     └──────────────────┘     └─────────────────┘

The server acts as a bridge between MCP-compatible AI assistants and the Replicate API, providing:

• Tool definitions that guide the AI on how to search and run models

• Smart response formatting for different output types

• Timeout handling for long-running predictions

License

MIT

Contributing

Contributions are welcome! Please open an issue or submit a pull request.

Credits

Built by Aaron Sherrill