Replicate MCP server

Read from the package source without running it. The installed server may list more.

• cancel_predictionwrite action

  Cancel an in-progress prediction. This stops the model execution and prevents further processing of the prediction.

• cancel_trainingwrite action

  Cancel an active or scheduled training session. Once cancelled, the training will no longer be available and participants will be notified of the cancellation.

• create_deploymentwrite action

  Create a new deployment to run a specific model version on designated hardware with auto-scaling configuration. The deployment will be immediately available for inference requests.

• create_deployment_predictionwrite action

  Create a prediction by running a model on a deployment with specified inputs. The request can optionally wait synchronously for results (up to 60 seconds) or return immediately with a prediction ID for asynchronous polling.

• create_filewrite action

  Upload a file with its content and optional metadata. The file is stored and can be referenced by its returned ID for use in other operations.

• create_modelwrite action

  Create a new model in your account. Each account is limited to 1,000 models; for iterative improvements, create new versions of an existing model instead.

• create_predictionwrite action

  Submit a prediction request to run a model with specified inputs. The request can optionally wait synchronously for results (up to 60 seconds) or return immediately with a prediction ID for asynchronous polling.

• create_prediction_for_official_modelwrite action

  Create a prediction by running an official model. The request can optionally wait synchronously for up to 60 seconds for the model to complete, or return immediately with a starting status for asynchronous polling.

• create_trainingwrite action

  Start a new training job for a model version. The training will create a new version of the model at the specified destination when complete. Since training can take several minutes or longer, use webhooks or polling to track completion.

• delete_deploymentwrite action

  Delete a deployment that has been offline and unused for at least 15 minutes. The operation returns a 204 status code on successful deletion.

• delete_filewrite action

  Permanently delete a file by its ID. Once deleted, the file resource will no longer be accessible and subsequent requests will return a 404 Not Found error.

• delete_modelwrite action

  Permanently delete a private model you own. The model must have no versions associated with it—delete all versions before attempting to delete the model itself.

• delete_model_versionwrite action

  Permanently delete a specific model version and all associated predictions and output files. Deletion is asynchronous and may take several minutes to complete.

• get_account

  Retrieve information about the authenticated account associated with the provided API token, including account type, username, display name, and GitHub URL.

• get_collection

  Retrieve a collection of models by its slug, including the collection metadata and all models within it.

• get_default_webhook_secret

  Retrieve the signing secret for the default webhook endpoint. Use this secret to verify that incoming webhook requests are authentically from Replicate.

• get_deployment

  Retrieve detailed information about a specific deployment, including its current release configuration, hardware settings, and scaling parameters.

• get_file

  Retrieve detailed metadata and information about a specific file by its ID.

• get_file_download

  Download a file using authenticated access credentials. Requires the file ID, owner information, and a cryptographically signed URL that includes an expiration timestamp.

• get_model

  Retrieve detailed metadata for a specific model, including its latest version, input/output schemas, and example prediction. Returns comprehensive information about the model's configuration, visibility, and usage statistics.

• get_model_readme

  Retrieve the README documentation for a specific model. Returns the README content as plain text in Markdown format.

• get_model_version

  Retrieve detailed metadata and schema information for a specific version of a model, including its OpenAPI schema that describes the model's inputs and outputs.

• get_prediction

  Retrieve the current state and results of a prediction, including its status, output, logs, and performance metrics.

• get_training

  Retrieve the current state and results of a training job, including status, metrics, logs, and output artifacts.

• list_available_hardware

  Retrieve a list of all available hardware accelerators and compute resources that can be used for running models on Replicate.

• list_collections

  Retrieve a paginated list of model collections available on Replicate, each containing curated groups of related models organized by use case or capability.

• list_deployments

  Retrieve a paginated list of all deployments associated with your account, sorted by most recent first. Each deployment includes its current release configuration with model, version, and hardware settings.

• list_files

  Retrieve a paginated list of all files created by the user or organization associated with the API token, sorted with the most recent file first.

• list_model_examples

  Retrieve all example predictions that were saved by the model author to demonstrate the model's capabilities. Use this to browse illustrative examples; for just the default example, use the get_model operation instead.

• list_model_versions

  Retrieve all versions of a model owned by a specific user or organization, sorted with the most recent version first. Returns paginated results including version metadata and OpenAPI schemas.

• list_models

  Retrieve a paginated list of publicly available models, optionally sorted by creation date or latest version update.

• list_predictions

  Retrieve a paginated list of all predictions created by your user or organization, including those from both API and web sources. Results are sorted with the most recent prediction first, returning up to 100 records per page.

• list_trainings

  Retrieve a paginated list of all trainings created by your user account or organization. Results include trainings from both API and web sources, sorted by most recent first, with 100 records per page.

• search_models

  Search for public models on Replicate using a text query. Returns a paginated list of models matching your search criteria.

• search_models_collections_and_docs

  Search across public models, collections, and documentation using a text query. Results include detailed metadata such as AI-generated descriptions, tags, and relevance scores to help identify the most suitable resources.

• update_deploymentwrite action

  Modify an existing deployment's configuration, including hardware SKU, scaling limits, and model version. Each update increments the deployment's release number.

• update_model_metadatawrite action

  Update metadata properties for an existing model, including description, documentation, and resource links. Only specified properties are updated; omitted properties remain unchanged.

• Code scan5 source files scanned25/25
• –Live reliabilityno gateway calls yet and no remote to proben/a
• –Tool poisoningtools not inspected (local package is not executed); not countedn/a
• Auth qualitylocal package, no credentials required12/15
• Maintenancelast push 138 days ago8/15
• Maintainer identityregistry namespace matches repository owner7/10

Safety score

80/100

higher than 60% of 48,255 servers

Tools

37

the typical server has 7

Tools that change things

41%

typical: 0%. Fewer write tools means less an agent can break

GitHub stars

31

more than 92% of 28,688 servers

Downloads a week

158

more than 71% of 12,191 servers with a package

Replicate: common questions

Is Replicate MCP server safe?

Mostly: it is graded B (80/100). Read the Replicate safety report

How do I install Replicate?

It runs on your machine. Copy the Claude Code, Cursor, VS Code or Claude Desktop config from the install section.

Does Replicate need an API key?

Not as far as the registry entry and our scan can tell: no credentials are declared or required.

Is Replicate maintained?

The last commit was 176 days ago (2026-05-12). The latest release is v1.0.2.

How does Replicate compare with other MCP servers?

safety score 80/100 (higher than 60% of 48,255 servers); tools 37 (the typical server has 7); tools that change things 41% (typical: 0%. Fewer write tools means less an agent can break). Full comparison.