Replicate MCP. Run ML Models, From Search to Output.

Replicate MCP Server

Connect your AI client directly to Replicate for thousands of open-source machine learning models. You don't need to set up local environments or manage GPU resources yourself; your agent handles it all on the backend. It lets you use complex ML predictions—like image generation or specialized LLMs—just by sending simple text commands.

Running and Monitoring Predictions

The core function is running model predictions. You call create_prediction when you want to start a new run; this requires you to supply the exact model version ID and all necessary input variables in JSON format. To keep tabs on what's happening, use get_prediction to check the current status or grab the final output of any prediction you started earlier.

If a process runs wild or you change your mind, you can immediately halt it using cancel_prediction, which kills a running model job by its unique ID.

Finding and Inspecting Models

If you need to find a model for a specific task, use search_models to scan the public catalog with keywords. If you know the general category, try list_collections to see curated groups of models—for example, 'Image-to-Text' or 'Audio Generation.' You can also pull a list of every available public model using list_models.

When your client finds a promising candidate model ID, it needs its specific requirements; run get_model to retrieve all the details, including the exact input schema and parameter rules for that single model. To see what's currently running or deployed within your organization, you can check out list_deployments, which shows your active models and their status.

Advanced Discovery and System Checks

For a deeper dive into available tools, use get_collection by providing a specific collection slug to fetch all the related models in that group. You can also see what GPU hardware options are available for running inferences on Replicate using list_hardware. To keep track of past work, list_predictions displays your full log of recent prediction history, including status updates and links to outputs.

If you need basic verification of your access, run get_account; this tool retrieves essential details about your authenticated Replicate account.

Putting It All Together

Your AI client can build a whole workflow using these tools. You start by running search_models for 'text-to-image,' then use get_model on the best result to confirm the required JSON structure, and finally execute create_prediction. If you want to make sure everything is working right before calling it, you can check your active deployments with list_deployments or see what collections are out there using list_collections.

This server gives your agent direct control over a massive library of open-source algorithms without needing any local setup. It's all about sending the right commands to get results.