An MCP server that serves knowledge from Markdown files — and nothing else

The idea is simple: a capsule is a connected folder of Markdown files. A manifest names the entry points and describes how the topics relate. When you ask a question, the LLM gets the map of the capsule and navigates it — reading files through a tool, following the references relevant to your question — until it has enough to answer. It cites what it used, and refuses when the capsule doesn't contain the answer.

Think of it as RAG with a hard wall and a sense of direction. Useful for assistants over internal documents, bounded support bots and expert knowledge surfaces where dumping the whole corpus isn't an option. Authored as connected Markdown, navigated by the model itself, leak-proof by construction.

The MCP server serves capsules to Claude Code, Cursor, Windsurf, Cline, Zed and Claude Desktop. It ships with a dependency-free stub model for offline tests and supports Anthropic, OpenAI, LM Studio and Ollama as navigators.