Manage Replicate ML Models via CLI
An MCP server exposing Replicate's model-hosting platform as tools for searching, running, and managing predictions - now unmaintained.
Why it matters
Integrate with Replicate's machine learning models through a command-line interface. Search, run, and manage predictions with ease.
Outcomes
What it gets done
Search and discover available machine learning models on Replicate.
Execute models with custom inputs and track prediction progress.
View and manage past predictions and generated outputs.
Install
Add it to your toolbox
Run in your project directory:
curl -fsSL https://spark.entire.vc/get/vb-replicate | bash Capabilities
Tools your agent gets
Search models using semantic search
View available models
Get detailed information about a specific model
View model collections
Get detailed information about a specific collection
Run a model with your input data
Run a model with your input data and wait for completion
Check the status of a prediction
Overview
Replicate MCP Server
This MCP server wraps Replicate's model-hosting API into tools an MCP client can call: model search and browsing, prediction creation/polling/cancellation, and generated-image viewing and cache management. It requires a Replicate API token configured in the client or as an environment variable. Use it to run Replicate models from an MCP client like Claude Desktop, Cursor, or Cline. It is explicitly marked not in active development - Replicate now offers an official MCP server, which is the safer choice for production use.
What it does
An MCP server that exposes Replicate's model-hosting platform as tools an MCP client can call directly: searching and browsing models and collections, creating and polling predictions, and viewing or managing generated images. The full tool set covers models (search_models, list_models, get_model, list_collections, get_collection), predictions (create_prediction, create_and_poll_prediction, get_prediction, cancel_prediction, list_predictions), and images (view_image, clear_image_cache, get_image_cache_stats).
When to use - and when NOT to
Use it when you want an MCP client such as Claude Desktop, Cursor, Cline, or Continue to run Replicate-hosted models - search, invoke, monitor, and cancel predictions - without leaving the chat interface, and you're comfortable using an experimental, unmaintained project. Do NOT rely on it for anything needing ongoing support: the repo is explicitly marked NOT IN ACTIVE DEVELOPMENT. Replicate now ships its own official MCP server, and this project's issues won't be addressed going forward, though contributions might still be folded in. Prefer Replicate's official server for production use; this one remains up for those who find it useful as-is or want to fork it.
Capabilities
Model tools for semantic search, browsing, and collection lookup; prediction tools to create a run (including a create-and-poll variant that waits until completion), check status, cancel a running prediction, and list recent predictions; and image tools to view generated images in the browser and inspect or clear the local image cache.
How to install
npm install -g mcp-replicate
Then add it to Claude Desktop's config (Developer section, Edit Config) with a Replicate API token from the account API-tokens page:
{
"mcpServers": {
"replicate": {
"command": "mcp-replicate",
"env": { "REPLICATE_API_TOKEN": "your_token_here" }
}
}
}
Alternative installs: build from source (git clone, npm install, npm run build, npm start) or run directly via npx mcp-replicate. The token can also be set as a plain REPLICATE_API_TOKEN environment variable for other MCP clients. Requires Node.js 18.0.0+ and TypeScript 5.0.0+; if tools don't appear, the README's troubleshooting steps cover checking the config, verifying the token, restarting the server, and reading server logs.
Who it's for
Developers already using an MCP client who want to run and manage Replicate model predictions from inside their chat workflow, and who are willing to accept an unmaintained, MIT-licensed project instead of Replicate's official supported server. It also suits anyone who wants to fork the source (git clone, npm install, npm run build/dev/start, with npm run lint and npm run format for code style) to extend it themselves, since the maintainers have said contributions might be folded in even though they won't actively address issues.
Source README
Replicate MCP Server
A Model Context Protocol server implementation for Replicate. Run Replicate models through a simple tool-based interface.
NOT IN ACTIVE DEVELOPMENT
This repo was an experiment in MCP tooling for Replicate. The company now offers an official MCP server. This repo will stay up for those who find it useful or want to fork it, but it's not in active development and issues won't be addressed. Contributions might be folded in but no promises. Enjoy at your own risk.
Quickstart
- Install the server:
npm install -g mcp-replicate
Get your Replicate API token:
- Go to Replicate API tokens page
- Create a new token if you don't have one
- Copy the token for the next step
Configure Claude Desktop:
- Open Claude Desktop Settings (⌘,)
- Select the "Developer" section in the sidebar
- Click "Edit Config" to open the configuration file
- Add the following configuration, replacing
your_token_herewith your actual Replicate API token:
{
"mcpServers": {
"replicate": {
"command": "mcp-replicate",
"env": {
"REPLICATE_API_TOKEN": "your_token_here"
}
}
}
}
- Start Claude Desktop. You should see a 🔨 hammer icon in the bottom right corner of new chat windows, indicating the tools are available.
(You can also use any other MCP client, such as Cursor, Cline, or Continue.)
Alternative Installation Methods
Install from source
git clone https://github.com/deepfates/mcp-replicate
cd mcp-replicate
npm install
npm run build
npm start
Run with npx
npx mcp-replicate
Features
Models
- Search models using semantic search
- Browse models and collections
- Get detailed model information and versions
Predictions
- Create predictions with text or structured input
- Track prediction status
- Cancel running predictions
- List your recent predictions
Image Handling
- View generated images in your browser
- Manage image cache for better performance
Configuration
The server needs a Replicate API token to work. You can get one at Replicate.
There are two ways to provide the token:
1. In Claude Desktop Config (Recommended)
Add it to your Claude Desktop configuration as shown in the Quickstart section:
{
"mcpServers": {
"replicate": {
"command": "mcp-replicate",
"env": {
"REPLICATE_API_TOKEN": "your_token_here"
}
}
}
}
2. As Environment Variable
Alternatively, you can set it as an environment variable if you're using another MCP client:
export REPLICATE_API_TOKEN=your_token_here
Available Tools
Model Tools
search_models: Find models using semantic searchlist_models: Browse available modelsget_model: Get details about a specific modellist_collections: Browse model collectionsget_collection: Get details about a specific collection
Prediction Tools
create_prediction: Run a model with your inputscreate_and_poll_prediction: Run a model with your inputs and wait until it's completedget_prediction: Check a prediction's statuscancel_prediction: Stop a running predictionlist_predictions: See your recent predictions
Image Tools
view_image: Open an image in your browserclear_image_cache: Clean up cached imagesget_image_cache_stats: Check cache usage
Troubleshooting
Server is running but tools aren't showing up
- Check that Claude Desktop is properly configured with the MCP server settings
- Ensure your Replicate API token is set correctly
- Try restarting both the server and Claude Desktop
- Check the server logs for any error messages
Tools are visible but not working
- Verify your Replicate API token is valid
- Check your internet connection
- Look for any error messages in the server output
Development
- Install dependencies:
npm install
- Start development server (with auto-reload):
npm run dev
- Check code style:
npm run lint
- Format code:
npm run format
Requirements
- Node.js >= 18.0.0
- TypeScript >= 5.0.0
- Claude Desktop for using the tools
FAQ
Common questions
Discussion
Questions & comments · 0
Sign In Sign in to leave a comment.