The generated service description is the clever part of the setup step: it reads what you dropped in and writes the summary the tool descriptions use, so the assistant's own advertisement of itself matches your API rather than a generic template. Because everything routes through Claude, the answers cost tokens on your key on every call.
An MCP server that reads a folder of your own documentation and turns it into two assistants: one that answers questions about the API, and one that turns a plain-language request into a query against it. Claude does the reasoning, and the prompts that steer it are files you can edit rather than behaviour baked into the code.
- `create-query` — hand it a `request` in plain language and get back a query written against your documented API
- `documentation` — ask a question and get the answer drawn from the documentation you supplied
- A `docs/` directory you fill with `.txt`, `.md` or `.json` files, which is the entire knowledge source
- Editable prompts under `prompts/` — `system-prompt.txt` for the main instruction, `tool-metadata.txt` for how the tools describe themselves, `query-metadata.txt` for query generation, and a `service-description.txt` generated during setup
An Anthropic API key, supplied as `ANTHROPIC_API_KEY`. Node.js 18 or higher; from a clone, `npm install`, `npm run build` and then `npm run setup`, which creates the `docs/` and `prompts/` directories, writes default prompt files and generates the service description. Add your documentation to `docs/` before use. The client entry points at `build/index.js` with the key in the `env` block, over stdio, and absolute paths are required.
One command plus a key — npx -y mcp-expert-server, then supply credentials