genmux MCP server: images from Claude, Cursor and OpenCode

The Model Context Protocol (MCP) is the standard way to hand an AI assistant a set of tools: a small server exposes functions, resources and prompt templates, and any MCP client — Claude Desktop, Claude Code, Cursor, OpenCode, Continue, Windsurf — can call them mid-conversation. @genmux/mcp is that server for genmux: once it is connected, "make me a 16:9 hero image for the pricing page, but tell me what it costs first" is a sentence you type to your assistant, not a script you write.

Install: one command

The server is an npm package, run on demand with npx. The only required setting is your key:

GENMUX_API_KEY=nb_live_... npx -y @genmux/mcp

Environment variables:

You do not run this by hand; the client starts it. What you configure is the client.

Client snippets

Claude Code and Claude Desktop

claude mcp add genmux -e GENMUX_API_KEY=nb_live_... -- npx -y @genmux/mcp

Cursor — .cursor/mcp.json in the project (or the global one from Cursor's settings):

{
  "mcpServers": {
    "genmux": {
      "command": "npx",
      "args": ["-y", "@genmux/mcp"],
      "env": { "GENMUX_API_KEY": "nb_live_..." }
    }
  }
}

OpenCode — an mcp block next to the provider block in opencode.json:

{
  "mcp": {
    "genmux": {
      "type": "local",
      "command": ["npx", "-y", "@genmux/mcp"],
      "environment": { "GENMUX_API_KEY": "nb_live_..." }
    }
  }
}

Continue — in config.yaml:

mcpServers:
  - name: genmux
    command: npx
    args: ["-y", "@genmux/mcp"]
    env:
      GENMUX_API_KEY: nb_live_...

Hosted endpoint, for clients without a shell

Some clients cannot spawn a local process — a browser-based assistant, a locked-down laptop, a team that would rather not distribute keys into config files. For those, the same server runs hosted at https://mcp.genmux.tech/mcp over Streamable HTTP, authenticated with the key as a Bearer token:

{
  "mcpServers": {
    "genmux": {
      "url": "https://mcp.genmux.tech/mcp",
      "headers": { "Authorization": "Bearer nb_live_..." }
    }
  }
}

In Claude's web and desktop apps this is a "custom connector" with that URL. The tool list is identical; the difference is where the key lives.

What the server exposes

Tools — what the assistant can call:

Resources — documents the assistant can read into context: genmux://docs/llms.txt, genmux://docs/openapi.yaml, genmux://models, and code snippets at genmux://snippets/python, /node, /curl, /openai-sdk.

Prompts — two templates: integrate-genmux (walks the assistant through adding the API to your codebase, using the snippets) and pick-a-model (asks the right questions and recommends a model and resolution).

There is deliberately no chat tool. Your assistant already is a chat model; if you want DeepSeek or Qwen3 in it, that is the OpenAI-compatible endpoint, not MCP — see the OpenCode or Continue posts.

A worked example

With the server connected in Claude Code, in a project that has a public/ folder:

You: I need a hero image for the pricing page — a clean isometric illustration of stacked coins turning into pixels, dark background, 16:9. Use nanobanana. Estimate the cost first and don't generate until I say so.

The assistant calls estimate_cost with {"model": "nanobanana", "resolution": "1K", "ratio": "16:9", "n": 1} and reports:

Assistant: One 1K image on Nano Banana is 2 credits (about $0.025 at the Scale rate); a 2K version would need Nano Banana Pro at 6 credits. Your balance is 1,184 credits. Generate at 1K?

You: Yes, save it to public/hero-pricing.png.

It calls generate_image with save: true and GENMUX_OUTPUT_DIR pointing at public/, waits for the job (Nano Banana at 1K takes a few seconds), and comes back with the file path, the job id and the actual charge from the job record. If you ask for three variants (n: 3) the estimate becomes 6 credits≈ $0.073 × 1K Nano Banana and the assistant should say so before spending it.

The behaviour to expect from a good client: it estimates on its own when a request is expensive, it uses list_models when you name a model loosely ("the Pro one"), and it reads genmux://snippets/node before writing integration code instead of guessing the request shape. The integrate-genmux prompt template makes that last habit explicit.

Why this beats a shell script

You could wrap curl calls in a script and have the assistant run it. The MCP server gives three things a script does not: typed tool schemas the model understands without documentation, estimate_cost as a first-class step (so the assistant asks before spending), and resources the model can read to write correct code. Failed jobs are refunded automatically, so an assistant that gets a prompt wrong costs nothing — but an assistant that generates ten 4K images when you meant one is the case estimate_cost exists for.

Getting started

npx -y @genmux/mcp with your key, one of the client snippets above, and your assistant can list models, quote a price and generate images from the same balance that pays for DeepSeek and Qwen3. Model details and per-resolution prices are on the Nano Banana page and the pricing page; the MCP page keeps the current tool list and client snippets, and the REST API the server wraps is documented in the docs.