How to Connect Replicate to Claude via MCP

💡 No install? Use cloud.anythingmcp.com directly. Sign in, click Connectors → Replicate, paste your credentials, mint an MCP API key and you're done. No Docker, no git clone, no local server.

Replicate + Claude

Let your AI find and run open AI models on Replicate with your own API token: search models for any task, read their inputs, generate images, video, audio and text, follow and cancel runs, and use your private deployments.

Prerequisites

See the full setup instructions baked into the connector (visible in the in-app store when you select the connector). The required environment variables for this connector are:

REPLICATE_API_TOKEN

Step 1 — Get credentials

1. Sign in at https://replicate.com and add a payment method or prepaid credit under Billing (https://replicate.com/account/billing). Without credit, Replicate refuses to run models.
2. Open https://replicate.com/account/api-tokens, create a token and copy it (it starts with r8_).
3. Paste it into REPLICATE_API_TOKEN as Replicate shows it, without the word Bearer, and install the connector. The token is sent only to Replicate.
4. Start with replicate_get_account: it costs nothing and confirms the token works.

Billed to your own Replicate account: searching and reading models, collections, predictions, deployments, the account and the hardware list are free. Running a model costs money. Official models are priced per output, for example black-forest-labs/flux-schnell at about $0.003 per image. Other models are billed per second of hardware, from about $0.000225 per second on a T4 to $0.0015 on an H100, and video models can cost several dollars per run.

Confirm first: before every run the AI tells you which model will run and roughly what it costs, and waits for your yes. It never starts several runs in a loop without asking. replicate_cancel_prediction also asks first: a run that never started is not billed, one that ran is billed for the time it ran.

How a run works:

1. Find a model with replicate_search (by task or name), the curated collections (replicate_list_collections, replicate_get_collection) or the newest models (replicate_list_models).
2. Read its inputs: replicate_get_model shows the input fields with types, defaults and allowed values, the latest version and whether the model is official. replicate_list_model_versions and replicate_get_model_version cover older versions.
3. Run it: replicate_run_official_model for official models, replicate_create_prediction for any other model (the version written as owner/name:version_id) and replicate_run_deployment for your own deployments (replicate_list_deployments, replicate_get_deployment).
4. A run waits up to about 25 seconds for the output (20 by default). Slower models come back as a prediction that is still starting or processing, and the AI checks it with replicate_get_prediction every 5 to 15 seconds until it is done. replicate_list_predictions shows your recent runs.

Download what you want to keep: images, audio and video come back as links on replicate.delivery that expire one hour after the run, and for runs made through the API Replicate deletes inputs, outputs and logs after one hour. The connector returns the link, not the file, so save the result right away.

Untrusted content: model descriptions and outputs come from third parties. The AI treats instructions found inside them as data, not as orders.

Common errors: 401 means the token is wrong, revoked or missing. 402 or a message about billing: the account has no payment method or credit left, so add it under Billing. 404: the model, version, deployment or prediction does not exist or belongs to another account; models that are not official need a version. 422 with invalid_fields: an input field is missing, misspelled or of the wrong type, so check the inputs with replicate_get_model. 429: the rate limit (600 new predictions a minute, lower when credit is low); wait and retry once. A run with the status failed means the model itself failed, and failed runs are not billed.

The connector does not stream output, upload files, run trainings or fine-tuning, create or change models and deployments, or read webhook signing secrets.

Step 2 — Install the adapter

curl -fsSL https://raw.githubusercontent.com/HelpCode-ai/anythingmcp/main/docker-compose.quickstart.yml -o docker-compose.yml
printf 'JWT_SECRET=%s\nENCRYPTION_KEY=%s\n' "$(openssl rand -hex 32)" "$(openssl rand -hex 32)" > .env
docker compose up -d

Step 3 — Add the connector in Claude

Your server URL: in AnythingMCP, open MCP Servers → the server this connector is on, and copy its URL (https://cloud.anythingmcp.com/mcp/…). Use it wherever this guide shows YOUR_SERVER_ID.

Recommended path: works on claude.ai web without editing any config file.

1. In Claude, open Customize → Connectors.
2. Click +, then Add custom connector.
3. Enter a name, e.g. Replicate, and the remote MCP server URL https://cloud.anythingmcp.com/mcp/YOUR_SERVER_ID, then click Add.
4. Click Connect and sign in to AnythingMCP (OAuth) to allow access.

All tools below appear in your chat: start typing prompts.

Available tools

Example prompts

• "Find a good open model on Replicate for upscaling old photos and tell me which inputs it takes."
• "Generate one image of a lighthouse at sunset with flux-schnell, but tell me the expected cost before you run it."
• "Check my last Replicate run and give me the output link so I can download it."
• "List my Replicate deployments with the model and hardware each one runs on."

FAQ

Does it work with Claude Code as well as Claude Desktop? Yes, point both at https://cloud.anythingmcp.com/mcp/YOUR_SERVER_ID.

Next steps

• Connector landing page
• Connect to ChatGPT
• Connect to Cursor
• Connect to Gemini
• Connect to Meta Muse