Replicate Extended MCP. Orchestrate ML Inference and Model Deployment.

This server connects your agent straight into Replicate's model infrastructure. You can run complex machine learning models—like stable diffusion or text generation—and manage their entire lifecycle using simple commands through your AI client.

Discovering Assets:
You can search the public catalog for ML tools using search_models based on keywords, and you'll find all available curated model groups by listing them with list_collections. To get the deep details on a specific collection or general model, use get_collection or get_model. You'll also check out your own user profile info using get_account, and if you need to verify incoming signatures for webhooks, call get_webhook_secret.

Running Predictions:
The core job is running models. Use create_prediction to kick off a new prediction run with specific input arguments and model versions. If you've already set up a dedicated deployment endpoint, you can run the task directly using create_deployment_prediction. Once that prediction is done, you check the output status and get the final data by calling get_prediction.

You also track all recent activity by listing past job runs with list_predictions.

Training Models:
When you need to fine-tune an existing base model on custom data, start that work using create_training. To keep tabs on how that training job is going, call get_training, which gives you the current status and progress metrics. You can also manage your own assets by calling create_model to register a new model within your account, or update its metadata like tags and descriptions with update_model.

If an existing model needs a tweak, use get_model_version to grab its detailed input schema and parameters.

Managing Deployments:
For production models, you've got two main tools. You can set up a brand new private deployment using create_deployment, letting you dictate custom autoscaling rules. Once that's live, running predictions through the dedicated endpoint is what create_deployment_prediction does. You can also see what hardware options are available for deployments by listing them with list_hardware.

Model Lifecycles:
You control everything from creation to deletion. After you create a model or version, you might need to clean up old stuff; use delete_model_version to remove specific historical versions of an asset. You can also get a full list of every version tied to a model with list_model_versions. If you're done with a prediction job that's still running, call cancel_prediction to stop it immediately.

Tracking and Viewing Info:
You'll use get_model for general info on any ML asset. To get granular details about model versions, you gotta hit up list_model_versions. You can also pull a list of all available curated collections using list_collections, which helps when you need to browse multiple assets.

This gives your agent the full toolkit: from searching for models with search_models to setting them live with dedicated deployments, managing their version history, and keeping tabs on every single prediction or training job.