Replicate MCP. Run ML Models from Natural Conversation

This server lets your AI client run thousands of open-source machine learning models—text-to-image, LLMs, audio, video—right from your chat window or IDE. You don't gotta leave your workflow to manage model runs; you just talk to it.

Discovering Models and Collections

You wanna find a tool? You can list every available ML model using list_models, which gives you the owner, run count, and general description for each thing out there. If you know what you're looking for, use search_models to narrow down results by keyword in the name or description across the whole catalog.

For curated groupings, check out list_collections to see model categories—like text-to-image or LLMs—by their slug. You can also get specific details on a collection using get_collection, which takes a collection's slug as input.

To dig into specific tools, you use get_model with the full owner/name slug to pull up all the metadata for one model. If that model needs an update or you wanna lock down a build for testing, you run get_model_versions, which lists every historical version of that model, so you get the exact version ID needed later.

To know what hardware you're running on, check out list_hardware; this shows all available GPU options and their pricing details for when you run inference workloads. You can also use get_account to validate your API token and see your current usage limits, which is smart to do before a big job.

Running Predictions and Managing Runs

When it's time to run something—whether it's generating an image or asking a language model a question—you kick off the process with create_prediction. This tool starts the ML model by taking both the model slug and a specific input object. It immediately returns the initial prediction ID, which you need for everything else.

You got that ID? You track what's going down using get_prediction; this checks the status of your run—is it starting, did it fail, or is it done?—and pulls out the final output URLs when it succeeds. If a run goes sideways or you change your mind, you can stop it completely with cancel_prediction by passing its unique ID, which changes the prediction's status to 'canceled'.

You also get a history of everything you’ve run using list_predictions, showing their ID, status, and when they were created.

How It Works When You Use It

Just subscribe this server and drop your Replicate API Token in. Then, tell your agent what you need—say, 'Generate a futuristic cityscape.' Your AI client handles the rest: it translates that casual request into the necessary sequence of tool calls—maybe calling list_collections first to find the right model category, then using get_model and finally hitting up create_prediction.

You manage every step of the model lifecycle without ever opening a website or leaving your chat window.