Docs Decomposer

English | 简体中文

TypeScript PDF document processor with a token-efficient MCP server.

• Extracts PDF text with PDF.js and builds a nested section tree from the table of contents and bookmarks.

• Exposes documents through MCP tools with paged outlines, section search, and capped section reads, so an agent never has to ingest a whole PDF at once.

• Includes a terminal browser for opening PDFs in docs/ and paging through long section previews.

• Supports absolute or allowed-root-relative paths, project root registration, and optional resources and prompts.

Quick start

npm install
npm test
npm run build
DOCS_DECOMPOSER_ROOTS="$PWD/docs" npm run mcp
npm run tui

Related MCP server: jdocmunch-mcp

MCP client configuration

This is the form used by the project's .mcp.json:

{
  "mcpServers": {
    "docs-decomposer": {
      "command": "sh",
      "args": ["-c", "exec npm --prefix \"$DOCS_DECOMPOSER_HOME\" run mcp"],
      "env": {
        "DOCS_DECOMPOSER_HOME": "/path/to/docs_Decomposer",
        "DOCS_DECOMPOSER_ROOTS": "/path/to/documents"
      }
    }
  }
}

DOCS_DECOMPOSER_HOME is expanded by sh to locate the project; DOCS_DECOMPOSER_ROOTS is the document directory the server is allowed to read.

A simpler form without the shell wrapper:

{
  "mcpServers": {
    "docs-decomposer": {
      "command": "npm",
      "args": ["--prefix", "/path/to/docs_Decomposer", "run", "mcp"],
      "env": {
        "DOCS_DECOMPOSER_ROOTS": "/path/to/documents"
      }
    }
  }
}

See USAGE.md for tool flows, pagination, resources, prompts, root safety, and the register_root workflow.

Commands

License

MIT