MCP Index
Developer ToolsFreeactiveReviewed by MCPIndex

Replicate MCP

Run and orchestrate Replicate-hosted AI models from an MCP client — model discovery, prediction execution, and asynchronous output handling.

View on GitHubView on NPMBrowse all tools

Looking for more MCP servers? Browse the full MCP tools directory or explore more tools in Developer Tools.

OverviewConfigSetupFAQRelated
Reviewed by MCPIndex

MCPIndex assessment

Replicate MCP is an inference orchestration layer, not a magic model button. Its strategic value is giving an agent access to a changing catalog of specialized models; its engineering cost is accepting asynchronous jobs, mutable aliases, temporary outputs, and per-run billing. Pin versions, persist prediction IDs, and enforce budgets before exposing it to an autonomous loop. With those controls, it is the most flexible media-generation primitive in this directory; without them, it is a credit card attached to a stochastic retry loop.

Quick overview

Replicate MCP exposes a model marketplace to an agent, but the important architectural fact is that inference is asynchronous infrastructure, not a function call. A prompt can launch a prediction that enters a queue, transitions through provisioning and execution, emits logs, and produces an output URL that may expire or disappear according to the model's retention policy. Agents that treat run-model as synchronous will either time out, duplicate expensive jobs, or lose the output after reporting success. The server's value is model discovery plus orchestration: Claude can inspect available models, choose one by task, submit a prediction with typed inputs, poll its status, and hand the resulting image, audio, video, or JSON back into the workflow. The hidden edge case is model version pinning. Replicate model names are mutable pointers; a new version can change input schemas, output formats, latency, and cost without changing the human-readable name. Production agents must use immutable version IDs, validate the input schema before submission, and persist prediction IDs for idempotent polling. The model catalog is the feature; version drift is the bill.

Best for

The teams and workflows that benefit most from this tool.

Agent-selected image, audio, and video generation

Prototyping specialized models without self-hosting GPUs

Multi-step media workflows with polling and post-processing

Comparing model versions under a common API

What this MCP server is best for

  • Replicate MCP is useful when you want to extend an AI assistant with real tools and live system access.
  • It helps move from chat-only answers to real actions such as reading data, managing systems, or retrieving current information.
  • It is best for developer and technical workflows where AI needs controlled access to external tools or services.

When to choose it

Choose Replicate MCP when you want an MCP server focused on Developer Tools and need tighter integration with your existing tools.

Good fit

replicateai-modelsimage-generationaudiovideoinference

Limitations

Things to watch before choosing this tool.

Predictions are asynchronous and require lifecycle handling

Model aliases can drift without input-contract stability

Output URLs may be temporary

Unbounded agent loops can create unexpected inference costs

Replicate MCP Configuration

Use the following configuration as a starting point for Claude Desktop or any compatible MCP client, then replace placeholder credentials with your own values.

claude_desktop_config.json
{
  "mcpServers": {
    "replicate": {
      "command": "npx",
      "args": ["-y", "replicate-mcp"],
      "env": {
        "REPLICATE_API_TOKEN": "r8_your_token"
      }
    }
  }
}

How to set up Replicate MCP

These setup steps cover the typical installation flow for this MCP server.

  1. 1

    Create a Replicate API token with the narrowest account permissions available and store it as REPLICATE_API_TOKEN; never place it in a prompt or committed config.

  2. 2

    Install and configure the server with npx -y replicate-mcp, then expose only the model families your workflow actually uses.

  3. 3

    Pro-Tip: The fatal flaw 90% of developers commit is calling a model by owner/name without pinning its immutable version. The alias can move to a new schema overnight, and the agent will submit the old input shape successfully enough to waste credits or fail asynchronously. Store the version ID beside every production workflow, validate inputs against that version, and reject unpinned model calls.

  4. 4

    Treat every prediction as a job: persist the prediction ID, poll with backoff, and make retries idempotent so a network timeout does not launch a second paid inference.

  5. 5

    Set application-level cost and concurrency limits; one recursive agent loop generating video can exhaust a budget before a human sees the first frame.

  6. 6

    Verify with a low-cost text or image model, record the version ID, prediction ID, status transitions, and final output URL before moving to expensive media models.

Compatibility

Supported environments for Replicate MCP.

Claude Desktop

Yes

Cursor

Yes

VS Code

Yes

Local stdio server using a Replicate API token. Pin immutable model versions and persist prediction IDs for reliable polling.

Frequently asked questions

Common questions for Replicate MCP.

The agent says the prediction succeeded but no output is available — no error appeared. Why?

The silent failure is asynchronous output lifecycle confusion. The prediction reached a terminal status, but the agent treated a submitted job as completed, or the output URL expired before retrieval. A second variant is a webhook or polling race where the status was read before output metadata was committed. The engineering fix is to require terminal-status polling, persist the prediction ID, fetch output immediately after success, and copy durable artifacts to your own storage instead of treating Replicate's URL as permanent.

Why must production workflows pin model versions?

A human-readable model alias can point to a new immutable version with different inputs, output types, performance, or pricing. Version pinning makes the inference contract reproducible and lets you compare changes deliberately.

Can it run image, audio, and video models?

Yes, provided the selected model's input schema and output type are supported by the client workflow. Media outputs can be large and URLs may be temporary, so download and persist them in durable storage.

How do I prevent duplicate paid predictions?

Use a client-side idempotency record keyed by workflow ID and model version, persist the prediction ID before polling, and retry status retrieval rather than resubmitting after a timeout.

Replicate MCP vs Competitors

FeatureReplicate MCPCompetitor
Model breadth✅ Marketplace access to many hosted model versions❌ One provider's fixed model catalog
Execution model✅ Asynchronous predictions with status and output lifecycle❌ Synchronous chat-completions request
Reproducibility❌ Requires explicit immutable version pinning✅ Provider model IDs often have stronger stability guarantees
Media modalities✅ Image, audio, video, and custom model outputs❌ Text-first APIs with narrower media support
Operational cost control❌ Agent must manage queues, retries, and budgets✅ Managed chat APIs simplify request accounting

Related Guides

How to Install MCP Servers

Cross-client installation guide for Claude Desktop, Cursor, and VS Code.

Claude Desktop MCP Setup

Beginner-friendly setup guide for Claude Desktop MCP.

Best Open Source MCP Tools on GitHub

Compare popular open source MCP tools by use case and setup difficulty.