Replicate MCP Server

A universal, cross-platform MCP server that gives any MCP client — Claude Desktop, the claude.ai web app (Connectors), Cursor, Cline / Continue in VS Code, or your own application — native access to the full Replicate catalog: image generation, video, music, speech, upscaling, segmentation, LLMs, voice cloning, 3D, lipsync — anything Replicate hosts.

It speaks both transports: stdio (the standard for local desktop/editor integrations) and HTTP/SSE (for remote/web clients and your own services).

Once connected to your MCP client, you can simply ask:

"Generate a cinematic shot of a lighthouse in a storm, 21:9" "Write a 30-second synthwave track" "Make a 5-second video of a paper airplane flying through a city" "Read this paragraph in a British male voice" "Upscale this image 4x"

…and the assistant calls the right Replicate model, waits for the result, and (on local/stdio setups) downloads the output to your machine.

Install

Fastest — one-click via Smithery (): open the server page and copy the Install snippet for your client (Claude, Cursor, VS Code, Cline, Windsurf, …). It wires the server in and prompts for your Replicate token. (Routes through Smithery's hosted gateway — easiest, zero local setup.)

Token-private — npm + stdio (your Replicate token never leaves your machine):

npx -y replicate-mcp-server

…or drop it into any MCP client's config (full walkthrough under Installation).

Claude Desktop one-file install: download the latest .mcpb from Releases and drag it onto Claude Desktop.

Already running (bring your own token): https://replicate-mcp.sena-labs.dev/mcp — HTTP/SSE, multi-tenant; pass your token in the x-replicate-api-token header.

Once connected, try the example prompts below.

Try it

Once your client is connected, ask in plain language — the assistant picks the right tool, waits for the result, and downloads the output locally:

"Generate a cinematic shot of a lighthouse in a storm, 21:9" "Which is the cheapest model for a 5-second video? Then use it." — recommend → generate "Make 4 logo variations of a fox, flat vector style" — batch (concurrent) "Transcribe this audio, summarise it, then read the summary aloud" — DAG pipeline "Upscale this image 4×" · "Write a 30-second synthwave track" · "Read this in a British male voice" "Kick off a Flux LoRA fine-tune on my dataset and tell me when it's done" — training

Discovery tools (replicate_recommend_model, replicate_search_models, replicate_estimate_cost) and the 5 built-in workflow prompts help it chain multi-step jobs without you naming models.

What's inside

36 tools and 66 curated models, designed to be both ergonomic for common cases and fully open-ended for everything else. Every generate/run tool waits for completion and auto-downloads the outputs to your machine — you ask once and get back local files, not a job id to babysit:

Curated generation tools

Voice, 3D, and avatar tools

Orchestration tools

Fine-tuning (trainings)

Deployments

Prediction management + cost

Generic / discovery tools

Outputs:

• Image / video / audio: downloaded to ~/Downloads/replicate-mcp/<model>/<prediction_id>/ (configurable). Local paths and original Replicate URLs are both returned. For images, the response also includes an inline base64 preview (sized to stay under Claude Desktop's 1 MB tool-result limit; larger images fall back to the URL embed) plus three embed snippets (<details>-wrapped iframe viewer with Save button, responsive <img>, or markdown image) so the chat client can render the result inline at full size.
• Text (LLM, vision, classifier): the model's reply is surfaced in text_output and printed at the top of the tool response so Claude can read it directly.

Editing an image you have on hand

The editing tools (replicate_upscale_image, replicate_inpaint, replicate_remove_background, replicate_vision, image-to-video, lipsync, 3D-from-image) all take a URL. To turn a file into a URL, use replicate_upload_file:

• You have a local path (Claude Desktop): replicate_upload_file({ file_path: "C:/Users/you/photo.jpg" }) → URL → pass to the editing tool. If an image is dragged/pasted into the chat, save it to disk first — the server cannot read chat attachments directly, and Claude Desktop cannot reproduce a large image's exact bytes as a tool argument.
• You have the bytes in memory (claude.ai web with a code container): read the uploaded file in the container, base64-encode it, then replicate_upload_file({ base64_data: "data:image/png;base64,..." }) → URL → editing tool. base64_data accepts a bare base64 string or a full data:<mime>;base64,... URI.
• You already have a URL: pass it straight to the editing tool — no upload needed.

Why this over the official Replicate MCP?

Replicate ships its own official MCP (replicate-mcp on npm, hosted at mcp.replicate.com). It's a great, free, always-up-to-date 1:1 bridge to the REST API — Stainless-generated from Replicate's OpenAPI spec, ~37 generic tools that mirror each endpoint (account, collections, deployments, files, hardware, models, predictions, trainings, webhooks), plus a --tools=dynamic mode and a Deno "Code Mode" sandbox.

This server is built for a different job. The official MCP is a thin, generic API bridge: to make an image you call create_predictions with a raw model id and raw input JSON, and the LLM has to look up each model's input schema itself; it returns starting and does not wait or download (you opt into Prefer: wait, capped at 60s, then poll manually). This server is a batteries-included creative studio: task-oriented tools with sane defaults, a curated model registry, built-in orchestration, and cost awareness — every generate/run tool waits for completion and auto-downloads the outputs locally.

In short: the official MCP is the right tool if you want exhaustive, always-current raw REST access (and a free hosted endpoint with OAuth). This server is the right tool if you want a curated, orchestrated, cost-aware creative workflow where you describe what you want and get finished files back — no schema lookups, no manual polling, no manual downloads. They're complementary: use the official one for full-surface API plumbing, use this one for getting creative work done.

Prerequisites

• Node.js ≥ 20 (uses native fetch and Web Streams)
• A Replicate account with an API token: https://replicate.com/account/api-tokens
• An MCP client — e.g. Claude Desktop (macOS/Windows/Linux), the claude.ai web app, Cursor, Cline / Continue (VS Code), or any custom MCP host. The walkthrough below uses Claude Desktop as the example; the same command / args / env block works in any client's MCP config. For remote/web clients, see HTTP / SSE transport.

Installation

1. Clone and build

git clone <this-repo> replicate-mcp-server
cd replicate-mcp-server
npm install
npm run build

This produces dist/index.js, which is the server entry point.

2. Get your Replicate API token

1. Go to https://replicate.com/account/api-tokens
2. Click Create token
3. Copy the token (starts with r8_...)

3. Connect it to your MCP client (Claude Desktop shown)

The MCP command / args / env block below is identical across clients — only the location of the config file differs (Cursor: ~/.cursor/mcp.json; Cline/Continue: the extension's MCP settings; custom hosts: your own launcher). For Claude Desktop, find your config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

Add the replicate entry (merge with anything already there):

{
  "mcpServers": {
    "replicate": {
      "command": "node",
      "args": ["/ABSOLUTE/PATH/TO/replicate-mcp-server/dist/index.js"],
      "env": {
        "REPLICATE_API_TOKEN": "r8_your_token_here"
      }
    }
  }
}

⚠️ Use the absolute path to dist/index.js. Tildes (~) and relative paths don't work in this config.

Optional: set a custom download directory:

"env": {
  "REPLICATE_API_TOKEN": "r8_...",
  "REPLICATE_DOWNLOAD_DIR": "/Users/you/my-replicate-outputs"
}

4. Restart Claude Desktop

Fully quit and reopen Claude Desktop. You should see "replicate" listed in the tools panel (the hammer icon at the bottom of the chat).

VS Code (native MCP)

VS Code has built-in MCP support. This repo ships a ready-to-use .vscode/mcp.json that runs the published npm package locally and prompts for your token (stored masked, never written to the file):

{
  "inputs": [
    { "type": "promptString", "id": "replicate_token", "description": "Replicate API token", "password": true }
  ],
  "servers": {
    "replicate": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "replicate-mcp-server"],
      "env": { "REPLICATE_API_TOKEN": "${input:replicate_token}" }
    }
  }
}

Open the file in VS Code and click Start above the replicate server (or run MCP: List Servers). VS Code asks for the token once.

Note on the Smithery "One-Click Install": it wires VS Code to Smithery's hosted gateway (mcp.smithery.run), which runs the server on Smithery's infrastructure and would route your personal Replicate token through it on every call. For a personal API token, prefer the local stdio config above — the token never leaves your machine.

Usage examples

Once connected, just talk to your assistant naturally (examples use Claude phrasing, but any MCP-capable assistant works):

Images

Generate an origami fox in a misty forest, 16:9 aspect ratio.

High-quality images with text

Make a logo for a coffee shop called "Crema" — use Recraft.

Video

Create a 5-second video of a paper airplane gliding through a neon-lit city. Use Kling Pro.

(Videos take 1–5 minutes. If the call times out, Claude will automatically poll with replicate_get_prediction.)

Music

Write a 20-second instrumental synthwave loop with a heavy bassline.

Songs with lyrics

Use the ACE-Step model to generate a sad indie-folk song with these lyrics: [...]

Speech

Read this paragraph in a British female voice: [...]

Chat with an LLM

Ask Llama 3 70B to explain quantum entanglement in two sentences.

Image understanding

Look at this photo [URL] and tell me what objects are in the foreground.

Upscale

Upscale this image 4x with Real-ESRGAN: [URL]

Background removal

Cut the background out of this product photo: [URL]

Anything else

Search Replicate for "speech to text", then transcribe this audio: [URL]

(Claude picks Whisper via replicate_search_models → replicate_run_model.)

How async predictions work

Image generation usually finishes in seconds. Video, long music, and some heavy models can take minutes. The server handles this transparently:

1. You call replicate_generate_video.
2. The server waits up to timeout_ms (default 5 minutes) by polling Replicate every 2 seconds.
3. If it finishes in time → you get URLs and local paths.
4. If it doesn't → you get back pending: true and a prediction_id. Claude can call replicate_get_prediction later to retrieve the result.

You can bump timeout_ms up to 30 minutes if you want to wait inline:

Generate a 10-second Hunyuan video. Wait up to 20 minutes.

Configuration

Costs

Replicate charges per second of compute, varying by model. Approximate costs:

• Flux Schnell image: ~$0.003
• Flux Dev image: ~$0.025
• Flux 1.1 Pro image: ~$0.04
• Kling 1.6 Pro 5-second video: ~$0.45
• MusicGen 30-second clip: ~$0.05
• Kokoro TTS: ~$0.001 per request
• Llama 3 70B chat: ~$0.001 per 1K tokens
• LLaVA 13B vision: ~$0.001 per image
• Real-ESRGAN 4x upscale: ~$0.003 per image

See https://replicate.com/pricing for current rates and your billing dashboard for usage.

Troubleshooting

"REPLICATE_API_TOKEN environment variable is not set" You haven't put the token in the env block of your Claude Desktop config. Double-check the JSON and restart Claude Desktop.

"Server disconnected" in Claude Desktop Usually a path problem. Make sure the path to dist/index.js is absolute and the file exists. Test it manually:

REPLICATE_API_TOKEN=r8_... node /ABSOLUTE/PATH/dist/index.js

You should see replicate-mcp-server v1.0.0 ready. API token detected. on stderr. Press Ctrl+C to quit.

A model isn't in the curated list Just use its full identifier: model: "stability-ai/sdxl" works the same way. Or use replicate_search_models to find one.

Need a specific version of a model Use the owner/name:version_hash form: model: "black-forest-labs/flux-schnell:bf53bdb93d739c9c915091cfa5f49ca662d11273a5eb30e7a2ec1939bcf27a00".

Downloaded files are missing Check REPLICATE_DOWNLOAD_DIR. Files are organised as <dir>/<sanitized-model>/<prediction-id>/output-N.<ext>.

Deploy as platform (v3.0+)

Beyond a single local stdio client, the server also runs as a multi-user platform — HTTP/SSE transport (for claude.ai web Connectors, remote editors, and custom apps), multi-token pool, webhook-driven async completion, Docker, npm distribution.

HTTP / SSE transport

Run the server as an HTTP service instead of stdio:

# Local-only, no auth — for development.
node dist/index.js --http --port 8088

# LAN-exposed with Bearer auth — behind a private network.
node dist/index.js --http --host 0.0.0.0 --port 8088 --api-key your-shared-secret

Clients POST JSON-RPC to http://host:port/mcp with:

Authorization: Bearer your-shared-secret
Content-Type: application/json
Accept: application/json, text/event-stream
Mcp-Session-Id: <set by server on initialize, echoed on follow-ups>

Health probe (no auth): GET /health → {"status": "ok"}.

npm install (when published)

npx replicate-mcp-server --http --port 8088 --api-key SECRET
# or as a Claude Desktop launcher:
npx replicate-mcp-server

Docker

docker build -t replicate-mcp-server .
docker run --rm \
  -e REPLICATE_API_TOKEN=r8_... \
  -e LOG_LEVEL=info \
  -p 8088:8088 \
  replicate-mcp-server

The default CMD starts in HTTP mode on 0.0.0.0:8088. Add --api-key via docker run ... replicate-mcp-server --http --api-key SECRET if you expose the port beyond a private network.

Multi-token round-robin pool

For team / multi-tenant deployments, give the server multiple Replicate accounts to spread rate-limit headroom:

REPLICATE_API_TOKEN_POOL=r8_account_a,r8_account_b,r8_account_c \
  node dist/index.js --http --port 8088

Each Replicate API call rotates through the pool. Falls back to single REPLICATE_API_TOKEN if only one configured.

Webhook receiver (event-driven completion)

If the server is reachable from the public internet, you can replace polling with webhook callbacks:

REPLICATE_WEBHOOK_PUBLIC_URL=https://your.domain/webhook \
REPLICATE_WEBHOOK_PORT=8089 \
  node dist/index.js --http --port 8088

Replicate POSTs prediction completion to the public URL; the server resolves the awaiting tool call without polling. Per-prediction random token authenticates each callback.

Smithery listing

This server is live on Smithery: smithery.ai/server/sena-labs/replicate-mcp-server — scanned (36 tools, 5 prompts, 2 resources) and one-click installable into Claude, Cursor, VS Code, Cline, Windsurf, and other MCP clients. smithery.yaml defines the install-time config UI (token, optional pool, log level, download dir).

claude.ai web Connector

For listing as a Connector inside the claude.ai web app:

1. Deploy the server publicly with HTTPS (Render / Fly.io / Cloudflare / bare VPS behind a TLS-terminating reverse proxy).
2. Register the MCP HTTP endpoint at https://console.anthropic.com/settings/connectors.
3. Users add it from claude.ai → Settings → Connectors → Add custom.

Anthropic's Connector review is manual — provide the OAuth flow / API key input form they require, plus the public /mcp URL.

Architecture

src/
├── index.ts        # MCP server, tool registration, response formatting (handler factory)
├── replicate.ts    # API client, polling, output extraction (URLs + text), file download with retry
├── schemas.ts      # Zod schemas for tool inputs (with .strict())
├── models.ts       # Curated model registry per category (8 categories)
└── constants.ts    # Shared constants (timeouts, paths, character limits)

test/
├── stdio-test.mjs       # End-to-end MCP handshake + tool registration test (response correlation)
└── unit/                # node:test unit suites for pure helpers
    ├── extract-urls.test.mjs
    ├── extract-texts.test.mjs
    ├── infer-filename.test.mjs
    └── sanitize.test.mjs

The server uses the stdio transport (standard for local Claude Desktop integrations) and the modern registerTool API of the MCP TypeScript SDK. All tool inputs are validated by Zod schemas with .strict() enforcement — no unknown parameters slip through.

The 8 curated generation tools share a single makeGenerationHandler factory; each tool only declares how it maps its specific input fields onto the Replicate request body. Output URLs are extracted by recursively walking the prediction's output field, which can be a string, an array, a nested object, or any combination. Text outputs (from LLM / vision / classifier models) are surfaced through the same walker so non-URL strings appear in the response. Files are streamed to disk using Node's stream/promises.pipeline() so multi-GB videos don't blow up memory. Downloads retry once on transient failures (network error or 5xx) with exponential backoff; 4xx errors fail fast.

Testing

npm run build
node --test test/unit/*.test.mjs    # 43 unit tests on pure helpers
node test/stdio-test.mjs            # End-to-end MCP handshake + tool/list + tool/call sanity check

Support

If this project is useful to you, consider supporting its development:

☕ ko-fi.com/senalabs

Trademarks & disclaimer

This is an independent, community-built project. It is not affiliated with, endorsed by, sponsored by, or officially connected to Replicate, Inc.

• "Replicate" and the Replicate logo are trademarks of Replicate, Inc. This server is an interoperability client for Replicate's public HTTP API; the name is used descriptively (nominative fair use) to indicate compatibility — not to imply origin or endorsement. For Replicate's own official MCP server, see replicate.com/docs/reference/mcp.
• Built on the Model Context Protocol (MCP), the open standard created by Anthropic.
• Model names referenced (FLUX, Recraft, Whisper, Llama, etc.) are trademarks of their respective owners and are used only to identify the models you can run.

You are responsible for complying with Replicate's Terms of Service and the license of each model you invoke. Bundled third-party open-source components and their licenses are listed in THIRD_PARTY_LICENSES.md. The banner/logo use the Inter typeface (SIL Open Font License 1.1).

License

MIT — see LICENSE.