Best Tools to Turn PDFs and Office Files into Markdown for AI Agents
Language models read text, not PDFs. To let an agent use a contract, a paper or a spreadsheet you first need a converter that preserves reading order, tables and headings, and that produces output small enough to fit in the context window.
The tools below range from MCP servers you add to Claude or Cursor to document-extraction engines for pipelines, and a desktop app that indexes local folders. Look at the formats you need, whether OCR is included, and whether anything is uploaded.
AgentIndex · Updated 2026-10-14
Quick picks
- anymd — Rust MCP server and CLI that converts PDF, Office, EPUB, HTML, images, audio and video to Markdown locally, with no API key.
- kreuzberg — Rust document extraction engine for 57+ formats with bindings for nine languages; pick it for pipelines rather than chat.
- markdownify-mcp — simple MCP server that converts files and web content to Markdown.
- kordoc — converts Korean HWP/HWPX, PDF and Office files to Markdown, with CLI and MCP.
- stashbase — desktop app that indexes local folders of documents and shares them with Claude and Codex over MCP; semantic search is optional and uses OpenAI or OpenRouter.
Side-by-side comparison
| Tool | What it is | Stars | License | Last push | Install |
|---|---|---|---|---|---|
| anymd | MIT Rust MCP server and CLI that turns PDF, Office, EPUB, HTML, images, audio and video into Markdown locally, with no API key | 1.0k | MIT | 2026-10-11 | npm install -g @sylphx/anymd |
| pdf-reader-mcp | MIT MCP server with a single read_pdf tool for text, images and metadata, with parallel page processing; renamed to anymd | 1.0k | MIT | 2026-10-11 | npx @sylphx/pdf-reader-mcp |
| kreuzberg | Rust document extraction engine for 57+ formats with bindings for nine languages | 9.4k | MIT | 2026-10-13 | pip install kreuzberg |
| markdownify-mcp | MCP server that converts files and web content to Markdown | 3.0k | n/a | 2026-09-25 | pnpm install |
| kordoc | Convert Korean HWP/HWPX, PDF and Office files to Markdown, with CLI and MCP | 2.4k | MIT | 2026-10-13 | npm install kordoc |
| Skill_Seekers | Turn documentation sites, GitHub repos and PDFs into AI skills | 15.1k | MIT | 2026-09-30 | pip install skill-seekers |
| stashbase | Apache-2.0 desktop app that indexes local folders of PDFs, Office files, images, audio and video and shares them with Claude and Codex over MCP | 1.0k | Apache-2.0 | 2026-10-13 | brew install --cask liliu-z/stashbase/stashbase |
Tools in this list with an in-depth AgentIndex review: 7/7
How to choose
- List your formats first: PDF only, Office files, EPUB, web pages, scanned images, or regional formats such as HWP.
- Check whether processing is local. Several tools here run on your machine with no upload; others call hosted services.
- For scans you need OCR. Check whether it is built in or needs extra models or tools.
- Look for page ranges, token budgets and citations. They matter when documents are longer than the context window.
- Quoted speed and accuracy numbers are the authors' own benchmarks; test on your own documents.
FAQ
How do I let Claude read a PDF through MCP?
Add a document MCP server such as anymd, pdf-reader-mcp or markdownify-mcp to your client, then ask the assistant to read a file path. Restrict the server to the folders it needs, because by default some can read any local file.
Is pdf-reader-mcp the same as anymd?
anymd lists itself as formerly pdf-reader-mcp, from the same authors, and adds more formats such as Office files, images, audio and video. New users should start with anymd.
Do these tools send my documents to a server?
It depends on the tool. anymd runs locally and says nothing is uploaded, while StashBase sends content to OpenAI or OpenRouter only if you turn on its optional semantic search. Check each project's README for what leaves your machine.
How we compiled this list
Tools are chosen by hand from the projects in our index, favouring those we have reviewed in depth. Stars, licence, last-push date and install command come from each repository and are refreshed with the site data. We read each project's README and docs; we have not run every tool ourselves, so test before you depend on one.