Add the official Replicate MCP server to Claude Desktop, Claude Code and Cursor, then search models and run image, video and text predictions without leaving your chat or editor. Includes exact commands, JSON config, token safety, costs and a zero-setup alternative on PicassoIA.
You can now ask Claude to find an image model on Replicate, run it with your prompt, and hand back the finished file without leaving the chat. That is the whole point of the Replicate MCP server. It plugs Replicate's HTTP API into any client that speaks the Model Context Protocol, so Claude Desktop, Claude Code and Cursor can search thousands of models and create predictions on your behalf.
The setup takes about five minutes. What slows people down are the small details: which URL to paste, which transport to pick, where the token goes, and what shows up on the bill afterward. Below you get the exact commands for each client, a token safety checklist, a look at the experimental code mode, and a zero-setup route for readers who only want finished images and videos.
Replicate hosts a large catalog of image, video, audio and text models behind a single API. The official MCP server wraps that API as tools, so an assistant can call it directly instead of you juggling curl commands and browser tabs. According to Replicate's docs, it supports all of the operations in Replicate's HTTP API: searching, listing and inspecting models, then creating and fetching predictions.
Say you type: make a 16:9 photo of a lighthouse at dawn. The client forwards your message to the language model, the model decides it needs an image tool, and it calls the server. The server relays the request to Replicate, which runs the chosen model on its GPUs and returns an output URL. The sequence usually looks like this:
💡 Tip: If you already know the model you want, name it in the prompt. That skips the search step and keeps results repeatable from one session to the next.
The same loop works for video and audio models. Ask for a five second clip and the assistant picks a video model, fills in the duration field, and polls until the run finishes. Video runs take longer than images, so expect a wait of a minute or more instead of a few seconds.
Replicate's MCP page names three tools, and the docs list the underlying operations:
| Tool or operation | What it does |
|---|---|
search_models | Finds models by task or description |
create_predictions | Starts a run with the inputs you supply |
list_hardware | Shows the hardware available for running models |
models.list, models.get | Lists models and returns one model's metadata |
predictions.get | Fetches the status and output of a run |
Hardware matters because it sets both speed and price, so list_hardware is worth a try before a big batch.
You can run the server in two ways, and the choice decides how authentication works.
| Hosted server | Local server | |
|---|---|---|
| Address or command | https://mcp.replicate.com | npx -y replicate-mcp |
| Sign in | Browser approval flow | REPLICATE_API_TOKEN environment variable |
| Runs where | On Replicate's side | On your machine, needs Node |
| Docs verdict | Recommended for most users | Alternative with manual token setup |
Choose the hosted server unless you have a reason to run a process on your own machine. It removes the token file from your setup entirely, and it is the path every example in the next sections uses first.
⚠️ Heads up: Older tutorials tell you to install
mcp-replicate. That was a community experiment, and its README says it is no longer in active development. The official server replaced it, so skip the old package.
The hosted server is the one Replicate recommends. Clients connect to the SSE endpoint, https://mcp.replicate.com/sse, and you finish with a web-based approval step where you paste a Replicate API token. Create one at replicate.com/account/api-tokens before you start and keep it on your clipboard.
The local option runs the same server as a process on your machine. Set REPLICATE_API_TOKEN, start it with npx -y replicate-mcp, and stop it with Ctrl+C. Most clients launch it for you from a config file like this:
{
"mcpServers": {
"replicate": {
"command": "npx",
"args": ["-y", "replicate-mcp"],
"env": {
"REPLICATE_API_TOKEN": "your-token-here"
}
}
}
}
VS Code with GitHub Copilot reads .vscode/mcp.json and uses a top-level servers object instead of mcpServers, so copy the block and rename that one field. If the server never shows up, check that Node is installed and that npx runs from a normal terminal.
Claude Desktop is the quickest route because the hosted server needs no config file at all.
https://mcp.replicate.com/sse.Prefer the local server? Put the JSON from the previous section into claude_desktop_config.json and restart the app.
Start with something that forces the full search, inspect and run loop:
Search Replicate for a fast photorealistic text-to-image model, show me its input options, then generate a 16:9 photo of a lighthouse at dawn.
Fast FLUX variants such as FLUX Schnell are typical candidates for a request like that, though the search decides what comes up first. If the answer arrives as a link instead of an inline picture, that is normal: the server returns the output URL, and the client decides how to display it.
Once that works, get specific. Name the aspect ratio, the lens, the light direction and the number of variations you want. Vague prompts waste runs, and every run is billed.
Claude Code takes one command, and it keeps the connection across projects.
Install the CLI if you do not have it yet, then add the server:
npm install -g @anthropic-ai/claude-code
claude mcp add replicate https://mcp.replicate.com/sse --transport sse --scope user
The --scope user flag stores the server in your user configuration, so every project you open gets access to it.
Launch claude, type /mcp, choose the Replicate entry, and a browser window opens for the same approval flow Claude Desktop uses. Once it reports as connected, ask for a run in plain language.
Because Claude Code works inside your repository, it can do more than show a link. Try this: Generate three hero image options for the landing page, then download each output into public/images. The download step matters more than it looks, and the costs section below explains why.
If the tools do not appear, run /mcp again and check the status line for the server. A server that shows as disconnected usually needs a fresh authentication, not a reinstall.
Cursor reads MCP servers from a JSON file, so this setup is a few lines of config.
Open Cursor Settings, go to Tools & Integrations and click New MCP Server. Cursor opens ~/.cursor/mcp.json, where you add the hosted server through the mcp-remote bridge:
{
"mcpServers": {
"replicate": {
"command": "npx",
"args": ["-y", "mcp-remote@latest", "https://mcp.replicate.com/sse"]
}
}
}
For a single repository, create .cursor/mcp.json in the project instead, and use the local-server JSON from earlier with your token in env. Save the file and the server should appear in the same settings panel, along with its tools.
From there, Cursor's agent can generate placeholder art, product shots or icons while it edits the component that needs them. That is the real payoff: the asset and the code change land in one conversation.
The same server works beyond these three:
.vscode/mcp.jsonThe pattern never changes: point the client at the hosted URL or launch replicate-mcp locally, then authenticate once.
Three things decide whether this setup stays pleasant: how you handle the token, what each run costs, and how far you push automation.
.cursor/mcp.json or .vscode/mcp.json to .gitignore whenever they hold a token.The MCP server is a front door to the same API, so every prediction is billed to your Replicate account like any other API call. The price depends on the model and the hardware it runs on, which is why list_hardware and a quick look at model pages help before a large batch.
A simple habit keeps the bill predictable. Draft with a fast, cheap model such as FLUX Schnell, generate four options, pick the best composition, and rerun only that one prompt on a higher-quality model. Tell the assistant this rule once at the start of a session and it will follow it for every request after that.
| Item | What happens |
|---|---|
| A prediction you run | Billed to your Replicate account |
| API output files | Removed after an hour by default |
| API input files and logs | Removed after an hour by default |
| Predictions made on the website | Kept indefinitely |
💡 Tip: Output URLs from API predictions disappear after about an hour. Ask the assistant to download or upload every result right away, especially in a long session.
Replicate also ships an experimental code mode. Instead of calling one tool at a time, the model writes TypeScript and runs it inside a Deno sandbox. That suits jobs that chain several predictions, such as generating six variations and keeping the best two. It needs Deno installed. Add it to Claude Code with:
claude mcp add "replicate-code-mode" --scope user --transport stdio -- npx -y replicate-mcp@alpha --tools=code
The @alpha tag tells you what to expect. Treat it as a testing ground, keep the regular server installed next to it, and check its output before you rely on it for production work.
If you only want the finished image or clip and not an afternoon of configuration, PicassoIA keeps hundreds of models on browsable pages with no tokens to paste. Pair it with an LLM such as Claude Sonnet 5 to draft prompts, then generate in the browser.
PicassoIA Image is the platform's own text-to-image model, and it is a fast place to test a prompt before you automate anything.
💡 Tip: Change one variable per run. When the result improves, you know which edit caused it.
PicassoIA also exposes a Replicate-style REST API, so the create, poll and fetch pattern you just set up will feel familiar:
https://api.picassoia.com/v1, authenticated with a Bearer token from your accountPOST /v1/models/{owner}/{name}/predictionsGET /v1/predictions/{id}POST /v1/predictions/{id}/cancelFour models are available through the API and the MCP connection: PicassoIA Image, PicassoIA Image Editor Pro, PicassoIA Video and Seedance 2.5 Lite. Each account gets 5 concurrent predictions, shared across tokens and MCP connections, and prompts can run up to 4,000 characters. API and MCP access depend on your plan, so check the pricing page before you build on it, and read the docs at picassoia.com/en/api.
| Replicate MCP server | PicassoIA API and MCP | |
|---|---|---|
| Catalog | Thousands of models | Four models |
| Sign in | Replicate API token | Bearer token from your PicassoIA account |
| Billed to | Your Replicate account | Your PicassoIA plan |
| Best fit | Wide model choice and experiments | A fast, fixed set for images and video |
Pick the route that matches your day. If you live in a terminal or an editor, connect the Replicate MCP server with the command for your client and run the lighthouse prompt. If you want a result in the next sixty seconds, open Picasso IA and test the same prompt there. Run both, compare the outputs, and keep whichever fits your workflow, or use each for what it does best.
Three prompts to try on PicassoIA Image, each short enough to paste:
Browse every model, from image and video generators to language models, at picassoia.com/en/all-models, and start creating your own images on Picasso IA today.
