docling-rs-mcp

A Model Context Protocol server that converts documents to Markdown for AI agents — PDF with layout analysis, table structure and OCR, DOCX, PPTX, XLSX, HTML, EPUB, e-mail, scans and 40-odd other formats. It is docling.rs, a Rust port of IBM's docling, and it serves MarkItDown's convert_to_markdown(uri) tool under the same name and contract: a markitdown-mcp setup switches by changing the command.

uvx docling-rs-mcp --root ~/Documents

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "docling": { "command": "uvx", "args": ["docling-rs-mcp", "--root", "/Users/me/Documents"] }
  }
}

Claude Code: claude mcp add docling -- uvx docling-rs-mcp --root ~/Documents.

Tools

PDF models

Office, HTML, Markdown and e-mail need no models. PDFs and images use ONNX models for layout, tables and OCR (~720 MB). Fetch them once:

uvx docling-rs-mcp --download-models

They land in the docling.rs cache ($DOCLING_RS_CACHE_DIR, else ~/.cache/docling.rs) and the server finds them from then on; --download-models=all adds the code/formula enrichment models (~740 MB more). Until then PDFs convert from their text layer, as MarkItDown's do.

macOS: PyPI has no macOS wheel of docling-rs yet, so uv builds it from source on first use — install a Rust toolchain (rustup) and run the --download-models line in a terminal once; the build is cached, and an MCP client then starts the server without waiting for it.

What a call may reach

Local paths and file: URIs are confined to --root (repeatable; the working directory by default, except in /, where it is required). http(s) URLs are fetched only with --allow-url-fetch, and private or loopback addresses stay blocked. --http serves Streamable HTTP on 127.0.0.1:3001 instead of stdio. --help lists every flag; the security model is in docs/SECURITY.md.

The same server ships in the docling-rs binary (docling-rs mcp) and the ghcr.io/docling-project/docling-rs image, models included; with the docling-rs Python package installed it is also python -m docling_rs.mcp. On NVIDIA GPUs, uv run --with docling-rs-cuda python -m docling_rs.mcp.

MIT licensed.