The README is explicit that `MODEL` is set in the environment and not passed per request, which means the advanced arguments — `steps`, `seed`, `guidance` — only do anything when the server was started on `flux.1-schnell`. Decide which model you want before you wire it in; switching means editing the config and restarting, not asking differently.
A Node server that exposes Baidu's iRAG image generation as a single tool. You describe the image, it calls Baidu, downloads the result and hands back base64 plus, if you want, a file on disk. Which model is used is a configuration decision rather than a per-call argument, so the client cannot switch models mid-conversation.
- `generate_image` takes a `prompt` and returns the image
- Optional `refer_image` to condition on a reference image URL
- `n` generates 1 to 4 images per call; `size` accepts 512x512, 768x768, 1024x768 or 1024x1024
- Two models chosen through the `MODEL` setting: `irag-1.0`, Baidu's own, and `flux.1-schnell`, which additionally accepts `steps`, `seed` and `guidance`
- `RESOURCE_MODE=local` saves files to disk and returns base64 and the path; `RESOURCE_MODE=url` returns only the URL and base64
- Automatic retries and logging, with the retry count set by `MAX_RETRIES` and the request timeout by `API_TIMEOUT`
A Baidu API key from the Baidu Cloud console with Qianfan ModelBuilder selected, shaped `bce-v3/ALTAK-****/****`, passed as `BAIDU_API_KEY`. Node.js 18.0.0 or newer. Clone, `npm install`, `npm run build`, then point your client at `node` with `dist/index.js`. Package `irag-mcp-server`, version 1.0.0. `BASE_PATH` sets where images land in local mode; leave it empty and they go to a folder on your desktop.