Replicate MCP server

What the model can do

Replicate (hosted AI models): the account's predictions and trainings with status, timings, compute seconds and errors (never their inputs, outputs or logs), its deployments with hardware and instance counts, the hardware on offer, and public models with the inputs they take. When asked, can run a model or a deployment (billed), cancel a run, start a fine-tune, create, resize or delete a deployment, create, describe or delete a model, and delete a model version.

Reads

• replicate_predictionsReplicate predictions (model runs) made by this account, from the API and the website, most recent first: model, version, status…
• replicate_predictionOne Replicate prediction by id: model, version, status, timings, compute seconds, error, deadline…
• replicate_trainingsReplicate trainings (fine-tunes) made by this account, most recent first, or one by training_id: base model, status, timings, compute seconds, error…
• replicate_deploymentsReplicate deployments of this account, or one by owner and name: the model and version released, release number, hardware…
• replicate_modelOne Replicate model by owner/name: description, visibility, whether it is official, run count, latest version, and the inputs that version takes…
• replicate_search_modelsSearch Replicate's public models by words (beta): owner/name, description, official or not, run count, tags.
• replicate_hardwareThe hardware Replicate offers for deployments and models: name and SKU (e.g. gpu-t4).

Acts, when you ask

These change something in Replicate. The model is told to use them only when you ask for exactly that, and each one is granted by name: per team on a shared connection, by the owner on a personal one.

• replicate_create_predictionRun a Replicate model once with the inputs given.
• replicate_run_deploymentRun a Replicate deployment once with the inputs given.
• replicate_cancel_predictionCancel a running Replicate prediction.
• replicate_create_trainingStart a Replicate training (fine-tune) of a model version, pushing the result as a new version of the destination model…
• replicate_cancel_trainingCancel a running Replicate training.
• replicate_create_deploymentCreate a Replicate deployment: a private endpoint running one model version on chosen hardware, scaling between min_instances and max_instances.
• replicate_update_deploymentChange a Replicate deployment's hardware, minimum or maximum instances, or the model version it serves.
• replicate_delete_deploymentDelete a Replicate deployment.
• replicate_create_modelCreate an empty Replicate model under this account, to push versions or trainings into.
• replicate_update_modelChange a Replicate model's description, readme, or GitHub, paper, weights or licence link.
• replicate_delete_modelDelete a Replicate model.
• replicate_delete_model_versionDelete a version of a Replicate model, and with it every prediction made with that version and their output files.

Connect it

First connect Replicate on Dashboard → Connections, then add Memtro to your client once. The Replicate tools appear alongside the tools of everything else you have connected; nothing is configured per connector.

claude plugin marketplace add memtro-org/claude-memtro
claude plugin install memtro@memtro --config api_key=sk-memtro-...

{
  "mcpServers": {
    "memtro": {
      "url": "https://platform.memtro.com/api/mcp",
      "headers": {
        "Authorization": "Bearer sk-memtro-..."
      }
    }
  }
}

{
  "mcp": {
    "memtro": {
      "type": "remote",
      "url": "https://platform.memtro.com/api/mcp",
      "headers": {
        "Authorization": "Bearer sk-memtro-..."
      }
    }
  }
}

Who may use it

The key identifies the person, and the server offers that person the tools they may use everywhere else in Memtro. A shared Replicate connection is used by the teams an administrator granted it to, tool by tool, on the Permissions page; a personal connection is used by its owner alone, and its tools that act stay off until the owner allows them. See Teams and permissions and Connectors.

Ask it like this

• › Which predictions failed today?
• › How many compute seconds did yesterday's runs use?
• › What hardware is the image deployment running on?
• › What inputs does this model take?
• › Scale the image deployment down to zero minimum instances

Questions

Is this an official Replicate MCP server?

No. It is Memtro's MCP server with Memtro's Replicate connector behind it: the connector reads Replicate through its API with the credentials you connect on Dashboard → Connections, and Replicate publishes nothing here. The same endpoint carries every other system you connect and Memtro's memory tools.

What does it cost?

Nothing on top of Memtro. One price per organisation by its number of users, from $50 a month for up to 10 users after a 7-day free trial, with every connector, the MCP server, jobs and memory included. You bring your own model keys.

Can it change anything in Replicate?

Only when you ask, and only what it has been granted: run a model (billed per second or per output by Replicate); run a deployment (billed, and may wake an instance); cancel a running prediction; start a training (fine-tune) of a model version into a destination model (billed); cancel a running training; create a deployment of a model version on chosen hardware (billed while instances run); change a deployment's hardware, instance counts or version; delete a deployment (only after 15 minutes offline); create a model (public or private); change a model's description, readme or links; delete a private model with no versions; delete a model version, with all its predictions and output files. Every other tool reads, searches or lists. An administrator grants each acting tool by name per team, and a personal connection acts only after its owner allows it.

Which clients work?

Any MCP client that speaks Streamable HTTP: Claude Code, Cursor, Claude Desktop, OpenCode, Codex, Zed, Windsurf, Cline and the Anthropic and OpenAI SDKs. There is no local process to run; the server is remote and authenticated with your key.

Other MCP servers · Engineering and operations

• GitHub MCP server
• AWS CloudWatch MCP server
• Sentry MCP server
• ServiceNow MCP server
• Okta MCP server
• Twilio MCP server