My Docs MCP Server

Overview

The My Docs MCP Server is a lightweight, TypeScript‑based service that turns any directory of Markdown files into an AI‑friendly knowledge base. By indexing the full text of every file with MiniSearch and applying Japanese tokenization via wakachigaki, it offers near‑real‑time, high‑accuracy search and retrieval capabilities over local documentation. This solves the common pain point of AI assistants being unable to locate relevant information within a project’s own docs, enabling developers to keep their knowledge in plain text while still exposing it to Claude or other MCP‑compliant clients.

At its core, the server exposes two MCP endpoints:

• (tool) – performs a full‑text query against the indexed corpus, returning file names, paths, and contextual snippets. It accepts a required string and an optional to cap the number of results. The integration of wakachigaki means Japanese queries are tokenized into natural “分かち書き” units, dramatically improving relevance for non‑English documentation.
• (resource) – retrieves the entire contents of a specified Markdown file. Clients can request this by providing an MCP resource URI, allowing seamless chaining from search results to full document view.

Developers benefit from a clear separation between data preparation (indexing) and consumption. The server can be launched in development mode with TypeScript or as a pre‑built Node.js binary, and it communicates over standard input/output so any MCP client—including Claude Desktop, Inspector, or custom tooling—can connect without network configuration. The integration with MCP’s tool and resource paradigms means a Claude workflow can search the docs, then pass the selected file path back to for a deeper dive, all within a single conversation.

Typical use cases include:

• On‑the‑fly documentation lookup in a large monorepo where developers need quick access to API docs, architecture diagrams, or setup guides.
• Knowledge‑base augmentation for product documentation sites that are maintained in Markdown; the server can surface relevant pages as Claude answers user questions.
• Testing and debugging of AI assistants that rely on external knowledge; Inspector can be used to manually verify tool outputs before deploying to production.

What sets this MCP server apart is its focus on Japanese text. By leveraging wakachigaki as a tokenizer for MiniSearch, search precision for Japanese documentation is substantially higher than generic tokenizers. Additionally, the use of a lightweight in‑memory search engine keeps latency low even for thousands of Markdown files, making it suitable for both local development and CI pipelines where quick feedback is essential. The server’s design—pure Node.js, no external services, and a simple stdio interface—ensures it can be dropped into any workflow with minimal friction.