kie-mcp

A Model Context Protocol (MCP) server for kie.ai — one API for the best AI models. Plug in your kie.ai API key and give Claude (or any MCP client) the ability to generate images, video, music, and text across 100+ models: Veo, Runway, Kling, Seedance, Flux, Nano Banana, Seedream, Ideogram, Qwen, Hailuo, Wan, Suno, GPT, Gemini, Claude, and more.

Built on kie.ai's unified jobs API plus its dedicated Veo / Runway / Suno / Flux Kontext / GPT‑4o endpoints, with a bundled catalog of every market model (exact model IDs + input schemas) so the assistant can discover and run anything.

• kie_generate — run any of 100+ image/video/music models end‑to‑end (submit + wait for the result URL).
• Model discovery — kie_list_models and kie_get_model_schema let the assistant find the right model and learn its exact inputs (no guesswork).
• Dedicated helpers for the big APIs with extra features: Veo (1080p/4K, extend), Runway (+ Aleph video‑to‑video), Suno (music, extend, lyrics), Flux Kontext, GPT‑4o images.
• kie_chat — chat with GPT, Gemini, and Claude text models.
• Utilities — credit balance, download links, and file upload (turn a local/base64 image into a hosted URL for image‑to‑video etc.).
• Robust polling, timeouts, and clean error messages. Your API key stays in your environment.

The bundled catalog currently covers 112 market models (41 image, 67 video, 4 music) — see src/catalog-data.ts.

• Node.js 18+
• A kie.ai API key — create one at https://kie.ai/api-key (sign in at https://kie.ai, open the API Keys section).

npx github:neomorrison/kie-mcp

(The first run builds automatically.)

git clone https://github.com/neomorrison/kie-mcp.git
cd kie-mcp
npm install        # builds automatically via the prepare script

The server entry point is dist/index.js.

The server reads your key from the KIE_API_KEY environment variable.

Edit your claude_desktop_config.json:

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

{
  "mcpServers": {
    "kie": {
      "command": "node",
      "args": ["/absolute/path/to/kie-mcp/dist/index.js"],
      "env": { "KIE_API_KEY": "your_kie_api_key_here" }
    }
  }
}

Or, without cloning:

{
  "mcpServers": {
    "kie": {
      "command": "npx",
      "args": ["-y", "github:neomorrison/kie-mcp"],
      "env": { "KIE_API_KEY": "your_kie_api_key_here" }
    }
  }
}

Restart Claude Desktop afterward.

# from a cloned + built checkout
claude mcp add kie --env KIE_API_KEY=your_kie_api_key_here -- node /absolute/path/to/kie-mcp/dist/index.js

# or via npx
claude mcp add kie --env KIE_API_KEY=your_kie_api_key_here -- npx -y github:neomorrison/kie-mcp

Run node dist/index.js (or npx -y github:neomorrison/kie-mcp) over stdio with KIE_API_KEY set in the environment.

kie.ai has hundreds of models. The assistant doesn't need to memorize them — it discovers them:

1. Find a model → kie_list_models (filter by category and search).
2. Learn its inputs → kie_get_model_schema (exact field names, types, defaults, allowed values).
3. Run it → kie_generate with the modelId and an input object. It submits the job and polls until the output URL(s) are ready.

Example (what the assistant does under the hood):

// 1) discover
kie_list_models { "category": "image", "search": "nano banana" }
// 2) schema
kie_get_model_schema { "model": "google/nano-banana" }
// 3) generate
kie_generate {
  "model": "google/nano-banana",
  "input": { "prompt": "a corgi astronaut on the moon, studio lighting", "output_format": "png" }
}

Most image jobs finish in seconds; video/music can take a few minutes. If a job exceeds the wait budget, the tool returns a taskId you can poll with kie_get_task (or the matching *_get tool).

See .env.example.

• Async by design. Generation is a job: create → poll → result. The *_generate / kie_generate tools do this for you; raise waitSeconds for long video/music renders, or set wait: false to return immediately and poll later.
• Credits. Each generation consumes kie.ai credits; check kie_get_credits.
• File lifetimes. Generated URLs typically expire after ~24 hours (some sooner). Download what you need, or use kie_get_download_url.
• Keep your key secret. It lives in your MCP client's environment, never in this repo. .env is git‑ignored.

npm install        # install deps (also builds via prepare)
npm run build      # compile TypeScript -> dist/
npm run typecheck  # type-check only
npm run watch      # rebuild on change

Source layout:

src/
  index.ts          server bootstrap (stdio)
  config.ts         env config
  client.ts         HTTP client (auth, timeouts, {code,msg,data} envelope)
  poll.ts           generic async-job poller
  catalog.ts        catalog types + search/lookup helpers
  catalog-data.ts   generated catalog of market models (ids + input schemas)
  util.ts           result/error helpers
  tools/            one module per tool group (market, chat, veo, runway, suno, flux, image4o, account)

src/catalog-data.ts is generated from the kie.ai docs index (https://docs.kie.ai/llms.txt) and each model's doc page. Re‑extract it when kie.ai adds models.

This is an unofficial, community-built integration and is not affiliated with or endorsed by kie.ai. Model availability, parameters, and pricing are controlled by kie.ai and may change; see the official docs.

MIT © neomorrison