# Replicate [Health: Active]

**Category:** 🗣️ Conversational AI  
**Repository:** https://github.com/pipeworx-io/mcp-replicate  
**GitHub Stars:** 0  
**npm Downloads (last month):** 6948  
**Views:** 3  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/replicate

## Description
Replicate MCP.

## Tools
Capabilities this server exposes over MCP:

- **list_models** — List all public models available on Replicate, paginated via cursor; returns owner, name, description, run count, and latest version id for each model.
- **model** — Fetch metadata for a specific Replicate model by owner and name, including description, visibility, run count, latest version, and GitHub URL.
- **model_versions** — List all published versions of a Replicate model by owner and name; returns version ids, creation timestamps, and input/output schema for each version.
- **model_version** — Fetch full metadata for a specific Replicate model version by owner, name, and version_id, including the complete input/output OpenAPI schema and Cog version.
- **search_models** — Search public Replicate models by keyword query; returns matching models with owner, name, description, and run count.
- **collections** — List all curated Replicate model collections (e.g. "text-to-image", "speech-to-text"); returns collection slug, name, and description.
- **collection** — Fetch models belonging to a specific Replicate collection by slug (e.g. "text-to-image"); returns the collection description and a list of model objects.
- **create_prediction** — Start a Replicate model run by submitting a version id and input object; returns a prediction id and status that can be polled with the prediction tool until output is ready.
- **prediction** — Poll the status and retrieve output for a Replicate prediction by id; returns status (starting/processing/succeeded/failed/canceled), output URLs, and logs.
- **list_predictions** — List your recent Replicate predictions in reverse chronological order, paginated via cursor; returns prediction id, model version, status, and creation timestamp.
- **cancel_prediction** — Cancel a prediction.
- **list_deployments** — List your Replicate model deployments (dedicated, auto-scaled endpoints), paginated via cursor; returns owner, name, current model version, and scaling config.
- **deployment** — Deployment detail.
- **create_deployment_prediction** — Run via a deployment.
- **account** — Fetch your Replicate account details including username, name, GitHub URL, and account type.
- **hardware** — List available Replicate hardware types (CPU, GPU SKUs) with their names and SKU identifiers; use to select hardware when creating a deployment.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "replicate": {
    "command": "npx",
    "args": ["-y","@pipeworx/mcp-replicate"],
    "env": {
      "PLATFORM_REPLICATE_KEY": ""
    }
  }
}
```

**Requires environment variables:** `PLATFORM_REPLICATE_KEY` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What Replicate MCP server does

The Replicate MCP server connects MCP clients to the Replicate API at `https://api.replicate.com/v1`. It gives agents access to public model discovery, model and version details, collections, predictions, deployments, account information, and available hardware types.

Model tools cover searching and browsing Replicate's catalog. An agent can list public models, retrieve a model, inspect its versions, search by query, and view collections or an individual collection. Prediction tools start runs, retrieve an individual prediction, list recent predictions, and cancel a running job.

The service also supports deployment-based runs. Agents can list deployments, inspect a deployment, and create a prediction through a deployment, which is intended for faster cold starts. The available models and their behavior remain determined by Replicate.

## How it works

A prediction request returns before the model necessarily finishes. After calling `create_prediction` or `create_deployment_prediction`, the client should poll `prediction(id)` until the status is `succeeded` or `failed`. Alternatively, the request can include a webhook URL and optional webhook event filters so Replicate can notify an external endpoint.

For LLM-style output, callers can set `stream=true`. The response provides a `urls.stream` Server-Sent Events endpoint for receiving the stream.

The Replicate MCP server is available in two forms. The hosted endpoint is exposed through the Pipeworx gateway and can be configured as an MCP URL. The standalone package runs locally over stdin and stdout and exposes only the Replicate pack's tools. The hosted pack endpoint also lists shared Pipeworx meta-tools, including routing and discovery tools; its `initialize` response is the authoritative source for the tools available on a given day.

## Setup and configuration

For hosted use, add `https://gateway.pipeworx.io/replicate/mcp` as the MCP server URL in a compatible client. The full Pipeworx gateway at `https://gateway.pipeworx.io/mcp` is another option; both URLs reach the same gateway, while the listed direct tools differ.

For local use, configure the client to run `npx -y @pipeworx/mcp-replicate`. The process communicates through MCP over stdin/stdout and does not require a Pipeworx account or gateway round trip. The package can also be started directly with the same `npx` command.

Authentication uses `PLATFORM_REPLICATE_KEY`. The gateway documentation also describes a BYO API key passed as `?_apiKey=…`.

## Tools and capabilities

The Replicate MCP server includes these capability groups:

- Models: `list_models`, `model`, `model_versions`, `model_version`, `search_models`, `collections`, and `collection`.
- Predictions: `create_prediction`, `prediction`, `list_predictions`, and `cancel_prediction`.
- Deployments: `list_deployments`, `deployment`, and `create_deployment_prediction`.
- Account: `account` and `hardware`.

The gateway also provides shared Pipeworx tools on the hosted pack endpoint, while the local package does not. Tools can additionally be called over HTTP through gateway routes such as `/v1/tools/replicate_list_models`.

## Limitations and notes

Prediction execution is asynchronous, so an agent must poll or provide a webhook rather than assuming that creation returns final output. Streaming requires handling the returned SSE URL. Hosted pack connections include extra shared tool schemas, which can increase the tool list and context compared with the standalone process. The standalone process does not provide Pipeworx-wide routing through `ask_pipeworx`.

_Full upstream README: https://allmcps.com/mcp/replicate/readme_