Documents in.
Markdown out.

An MCP server that converts local documents to clean, token-efficient Markdown. 40-80% fewer tokens, better LLM reasoning.

$ npx distill-mcp-server click to copy
Works with Claude Desktop, Claude Code, Cursor, Windsurf, and any MCP client

Two modes. One tool.

Choose based on what you need. Switch anytime.

Lightweight

Python only

Requires: Python 3 + pip install distill-core

  • DOCX, DOC, ODT
  • XLSX, XLSM, CSV
  • PPTX, PPT
  • PDF (native text)
  • HTML
  • ~2 min setup
Full

Docker powered

Requires: Docker + Distill service

  • Everything in lightweight
  • Scanned PDFs (OCR)
  • Audio: MP3, WAV, M4A, FLAC, OGG
  • EPUB, JSON, SQL, WSDL
  • Quality score (0.0 - 1.0)
  • ~10 min setup

Supported formats

Validate before conversion. Unsupported files never leave your machine.

Category Extensions Lightweight Full
Word .docx .doc .odt ✓ ✓
Excel .xlsx .xlsm .csv ✓ ✓
PowerPoint .pptx .ppt ✓ ✓
PDF (native) .pdf ✓ ✓
PDF (scanned) .pdf - ✓
HTML .html .htm ✓ ✓
Audio .mp3 .wav .m4a .flac .ogg - ✓
EPUB .epub - ✓
Data .json .sql .wsdl .wsd - ✓

How it works

Three steps. No uploads. Everything stays local.

01

Give a file path

Tell your MCP client to convert a local file. Provide the absolute path - no need to upload anything.

02

Distill converts

The server sends the file to Distill (locally via Python or Docker). Clean Markdown comes back.

03

Cached & ready

The Markdown is cached locally and returned to the LLM. Repeated conversions are instant.

Quick setup

Add to your MCP client config. Here's Claude Desktop as an example.

{ "mcpServers": { "distill-mcp-server": { "command": "npx", "args": ["-y", "distill-mcp-server"], "env": { "DISTILL_MCP_CONFIG": "{\"mode\":\"lightweight\",\"python_path\":\"python3\"}" } } } } // Windows: change python_path to "py" // Full mode: change mode to "full" and add distill_url

Full setup guides: Lightweight · Full (Docker)

Privacy first

All processing happens on your machine. Lightweight mode uses a local Python library. Full mode runs Docker locally. No data is sent to external services.