What the assistant can look up is exactly the set of llms.txt URLs you pass at launch, which makes it predictable: add the libraries your project depends on and it will read their current pages, and it will not wander further. Adding a library later means editing the client entry rather than anything inside the server.
An MCP server that serves documentation through llms.txt. An llms.txt file is published by a site the way robots.txt is, listing its URLs with descriptions — this reads those lists so an agent can pull the current page rather than answering from training data.
- Documentation sources named at launch, each as a label and an llms.txt URL — the README's example pairs `LangGraph:https://langchain-ai.github.io/langgraph/llms.txt` with `LangChain:https://python.langchain.com/llms.txt`
- Several sources in one server, so an assistant can reach across the libraries a project actually uses
- The URL breakdown and descriptions that llms.txt carries, which is what lets the right page be chosen before it is fetched
- A transport and address you set on the command line — `--transport stdio`, with `--port` and `--host` alongside
`uv` on the machine and the repository available locally; the documented client entry runs `uvx --from <path to the project> mcpdoc` with the `--urls` list. The README's own instructions for filling in that entry are practical: take the command path from `which uv` and the `--from` path from `pwd` in the project folder. No key is involved — the llms.txt files it reads are public. The project is `mcpdoc` 0.0.8.
One command — uvx mcpdoc