Markitdown Rs MCP Server

markitdown-rs

markitdown-rs converts PDF, Word, PowerPoint, Excel, HTML, EPUB, Outlook and web pages to Markdown for LLMs and RAG pipelines. It is Microsoft's markitdown CLI rewritten in Rust as one 4.5 MB binary: the same command and the same Markdown, byte for byte, without a Python environment. From the command line it is usually hundreds of times faster (median 408× over 231 real-world files, about 15× on the largest PDFs and spreadsheets); in process, 4–50× faster.

brew install majiayu000/tap/markitdown-rs   # or: cargo install markitdown-rs
markitdown report.pdf > report.md

Prebuilt binaries for macOS (Apple silicon and Intel) and Linux x86_64 are on the releases page. It also runs as an MCP server for AI agents (markitdown mcp, stdio or HTTP) and has a Python module with markitdown's API, plugins included.

Example

A PDF form becomes a Markdown table, exactly as markitdown renders it (tests/fixtures/pdf/gen-form-mixed.pdf):

$ markitdown order.pdf
Purchase order summary
| Item     | Qty | Unit price | Total |
| -------- | --- | ---------- | ----- |
| Widget 1 | 1   | 1.50       | 1.50  |
| Widget 2 | 2   | 3.00       | 6.00  |
| Widget 3 | 3   | 4.50       | 13.50 |
...
Thank you for your order.

URLs work too, with markitdown's Wikipedia and YouTube handling:

$ markitdown https://en.wikipedia.org/wiki/Markdown | head -3
# Markdown

Plain text markup language

markitdown-rs vs. the alternatives

Full feature-by-feature table, with how each number was measured: markitdown vs markitdown-rs.

Why

markitdown is the de facto "documents to Markdown for LLMs" tool, but using it from a shell, CI job or non-Python service means shipping a Python environment with mammoth, pdfminer.six, pdfplumber, pandas, openpyxl, BeautifulSoup, python-pptx, magika and onnxruntime. Its most requested features include brew and binary installs (#1179, #25) and lighter dependencies (#1234).

markitdown-rs targets output compatibility. It does not try to produce "better" Markdown: it reproduces markitdown's output, quirks included (pipes in table cells, Heading 7–9 as plain text, pdfminer's (cid:N) placeholders, pandas' NaN cells), so you can swap the binary without changing anything downstream.

Compatibility

Every fixture is converted by both implementations and compared byte for byte in cargo test against Python markitdown 0.1.9 (see Parity testing): 408 files, the 6,705 inputs of the html5lib tokenizer suite through the HTML converter, and 13 pages converted as if fetched from Wikipedia, Bing and YouTube URLs.

Sources are files and stdin as in markitdown, plus http:// / https:// URLs (with markitdown's Accept header and stream detection from the response), file: and data: URIs.

What each converter covers:

• DOCX follows mammoth 1.13: headings, bold/italic/underline/strike, lists, tables with merged and nested cells, hyperlinks (including linked pictures), footnotes and endnotes, images, equations (OMML to LaTeX), text boxes, checkboxes and tracked changes.
• PPTX follows python-pptx: titles, text in reading order, groups, tables, pictures, charts (bar, line, pie, area, radar, scatter, bubble) and notes.
• PDF is a port of pdfminer.six and the parts of pdfplumber markitdown uses: xref tables and streams (with the fallback for broken files), object streams, Flate/LZW/ASCII85/ASCIIHex/RunLength with predictors, RC4 and AES encryption (empty user password), Type1/TrueType/Type3/CID fonts with ToUnicode maps, the standard 14 fonts' metrics, embedded TrueType cmaps and the predefined CJK CMaps, form XObjects, pdfminer's layout analysis, and markitdown's borderless-table detection on pdfplumber word positions.
• XLSX/XLS follow pandas' read_excel (openpyxl/xlrd value conversion, dtype inference, to_html), one table per sheet.
• HTML ports CPython's html.parser, BeautifulSoup's tree builder and markdownify with markitdown's overrides; Wikipedia, Bing and YouTube pages get their dedicated converters when the source URL says so.
• MSG reads the MAPI properties from the OLE container (code pages included); EPUB walks the OPF spine; ZIP converts every member.
• Images and audio print exiftool metadata when exiftool is installed, as markitdown does.
• YouTube pages get the video's transcript (### Transcript), fetched the way markitdown's optional youtube-transcript-api does.

Known differences:

• PDF text-box order on rare layouts. pdfminer breaks ties between equal box distances by id(), a CPython memory address. markitdown-rs uses creation order instead. Of 313 PDFs tried (all of DP-Bench, arXiv papers, a 10-K, CJK manuals, other projects' test files) 308 are identical, one polyglot file is rejected by both, and 4 differ only in where a tied box lands (they match a pdfminer patched to use creation order).
• Content sniffing uses rules instead of magika's model, so files without an extension (or with a wrong one) can be routed differently; charset detection uses chardetng instead of charset_normalizer, which disagrees on short legacy 8-bit text and on Outlook messages that declare the wrong code page (tests/known-diffs.txt).
• Style maps embedded by mammoth (a mammoth/style-map part) are ignored by the Rust converter; the Python module applies them when given a style_map (see Python).
• Deeply nested input. HTML nested deeper than Python's recursion limit allows (493 elements) falls back to plain text exactly where markitdown does, with the same warning, however deep it goes. Word tables nested 5000 deep make Python fail; markitdown-rs stops at 512 levels of XML nesting with an error.
• UTF-16 HTML with a byte order mark is decoded by its BOM. markitdown decodes UTF-16LE but turns UTF-16BE pages into mojibake.
• Wikipedia answers Python's default python-requests User-Agent with HTTP 403; markitdown-rs sends its own and gets the page.

The command line does not do what markitdown leaves to optional Python packages or cloud services: audio transcription (markitdown's speech_recognition calls Google's unofficial speech API with a key built into that library, and needs ffmpeg), LLM image captions, Azure Document Intelligence / Content Understanding (-d, --use-cu) and plugins (-p). The Python module does all of these, and its own markitdown command has the original's full set of flags. Like markitdown, legacy .doc and .ppt files are not converted.

Install

With Homebrew (macOS, Linux x86_64):

brew install majiayu000/tap/markitdown-rs

From crates.io (Rust 1.88+):

cargo install markitdown-rs

Prebuilt binaries for macOS (arm64, x86_64) and Linux (x86_64) are attached to each release, with SHA-256 sums:

curl -L https://github.com/majiayu000/markitdown-rs/releases/download/v0.2.1/markitdown-v0.2.1-aarch64-apple-darwin.tar.gz | tar xz
./markitdown-v0.2.1-aarch64-apple-darwin/markitdown --version

From a checkout: cargo install --path crates/markitdown-rs. All of these install a binary named markitdown. The PyPI package is also available: python -m pip install "markitdown-rs[mcp]" in a fresh Python 3.10+ environment.

Usage

The flags mirror the Python CLI:

markitdown [filename] [-o OUTPUT] [-x EXTENSION] [-m MIME_TYPE] [-c CHARSET] [--keep-data-uris]

markitdown report.pdf                   # Markdown on stdout
markitdown book.xlsx -o book.md         # write to a file (no trailing newline, like Python)
markitdown https://example.com/a.docx   # fetch a URL
cat report.docx | markitdown -x docx    # stdin, with an extension hint
markitdown --keep-data-uris report.docx # keep images as full base64 data URIs

As in markitdown, the format comes from the hints (extension, MIME type, charset) and from the file contents, and converters are tried in markitdown's order until one succeeds.

Limits that markitdown does not have, so that hostile files fail cleanly instead of exhausting memory or hanging (environment variables, also Options fields in the library):

As a library (markitdown-core; URL loading is the optional fetch feature):

use markitdown_core::{StreamInfo, convert};
let data = std::fs::read("report.pdf")?;
let md = convert(&data, &StreamInfo::from_path("report.pdf"), &Default::default())?;

MCP server

markitdown mcp serves the Model Context Protocol on stdio with the same tool as microsoft/markitdown-mcp: convert_to_markdown(uri) for http:, https:, file: and data: URIs (local paths work too). Logs go to stderr, so stdout carries only JSON-RPC.

Claude Code:

claude mcp add markitdown -- markitdown mcp

Claude Desktop (claude_desktop_config.json) and Cursor (~/.cursor/mcp.json, or .cursor/mcp.json in a project) use the same block:

{
  "mcpServers": {
    "markitdown": { "command": "markitdown", "args": ["mcp"] }
  }
}

Desktop apps may not see your shell's PATH; if the server does not start, use the full path from which markitdown (e.g. ~/.cargo/bin/markitdown) as command.

Tool name, input and output schemas, result shape and error messages match markitdown-mcp 0.0.1a8; serverInfo.version is markitdown-rs's version, and argument validation errors leave out pydantic's details.

Over HTTP, as markitdown-mcp --http:

markitdown mcp --http                      # http://127.0.0.1:3001
markitdown mcp --http --host 127.0.0.1 --port 8080

It serves Streamable HTTP at /mcp (stateless, JSON responses) and the SSE transport at /sse (messages to /messages/), checks Host and Origin against DNS rebinding when bound to localhost, and warns when bound to another interface: the server has no authentication.

markitdown plugins are Python code, so this server cannot load them (MARKITDOWN_ENABLE_PLUGINS has no effect, and it says so). For an MCP server with plugins such as markitdown-ocr, use the Python package: pip install 'markitdown-rs[mcp]' installs markitdown-mcp 0.0.1a8's server on the same converters, as markitdown-mcp and markitdown mcp with its flags (--http, --host, --port, --sse), honouring MARKITDOWN_ENABLE_PLUGINS=true.

Python

crates/markitdown-py builds a Python module with markitdown's API on top of the same converters (PyO3 + maturin). Install from PyPI with python -m pip install "markitdown-rs[mcp]", or build from source with maturin (crates/markitdown-py/README.md). Its import name is markitdown, the same as Microsoft's package, so it cannot be installed in the same environment as markitdown from PyPI: use one or the other.

The wheel also installs markitdown's command, markitdown (and python -m markitdown): 0.1.9's own __main__.py on these converters, with all of its flags: -p / --list-plugins, -d / -e, --use-cu and --cu-*, and audio transcription with markitdown-rs[audio-transcription]. So there are two markitdown commands with the same output where they overlap:

If both are installed, whichever comes first on PATH runs (which -a markitdown lists them). To pick explicitly, run python -m markitdown for the Python one, or the binary's full path (e.g. ~/.cargo/bin/markitdown) for the Rust one.

from markitdown import MarkItDown

result = MarkItDown().convert("report.pdf")   # path, URL, file:/data: URI, stream, requests.Response
print(result.markdown)                        # or result.text_content

The module follows markitdown 0.1.9's own design: the built-in converters are DocumentConverter classes in markitdown.converters, each backed by the Rust converter of the same name, tried in markitdown's priority order. So these work as in markitdown:

• convert, convert_local, convert_stream, convert_uri, convert_response, StreamInfo, the exception classes and result.title; output is byte-identical to the CLI.
• Custom converters: register_converter(converter, priority=...), PRIORITY_SPECIFIC_FILE_FORMAT / PRIORITY_GENERIC_FILE_FORMAT, subclasses of the built-ins (with the _image_to_html hooks of the DOCX, PPTX and XLSX converters), enable_builtins=False.
• Plugins (enable_plugins=True, markitdown.plugin entry points); the markitdown-ocr plugin gives the same output as with markitdown.
• llm_client / llm_model / llm_prompt for image descriptions and slide picture captions; requests_session; exiftool_path; docintel_* and cu_* (markitdown's Azure converters, which need the Azure SDKs).
• style_map: as in markitdown, the document goes through mammoth with the caller's and the embedded style map, then through the HTML converter (markitdown-rs[docx] installs mammoth).
• Audio transcription when speech_recognition is installed and YouTube transcripts when youtube-transcript-api is, as in markitdown. The wheel has markitdown's extras for them and for Azure (markitdown-rs[audio-transcription], [youtube-transcription], [az-doc-intel], [az-content-understanding], [all]); see crates/markitdown-py/README.md.

Installing plugins. markitdown's plugins (markitdown-ocr and others) depend on Microsoft's markitdown package, so a plain pip install markitdown-ocr also installs it, and it overwrites this package's markitdown module: import markitdown then runs Microsoft's code (tens of times slower) and uninstalling either package removes shared files. markitdown-rs warns about this, with the fix, when markitdown is imported (or the markitdown command runs), and also when another package's uninstall has left its markitdown module missing. Install plugins without that dependency instead:

pip install 'markitdown-rs[ocr]'          # markitdown-ocr's other dependencies
pip install --no-deps markitdown-ocr

With uv, drop the dependency for the whole project in pyproject.toml:

[tool.uv]
override-dependencies = ["markitdown; sys_platform == 'never'"]

or for one install: uv pip install markitdown-rs markitdown-ocr --override overrides.txt with that line in overrides.txt. If the warning appears, pip uninstall -y markitdown markitdown-rs, then install as above.

Differences: unknown keyword arguments are ignored as in markitdown but named in a warning (unless plugins are enabled, since plugins read their own), HTTP failures without a requests_session raise markitdown.HTTPError (an OSError) instead of requests.HTTPError, and the ZIP converter converts members with the built-in converters only. crates/markitdown-py/tests/test_api.py runs against both implementations.

Benchmarks

On 231 real-world files that are not part of this repository's fixtures (234 Apache Tika test documents and 10 arXiv papers, all those both implementations convert), measured on 2026-11-05 with markitdown-rs 0.2.0 on an Apple Silicon Mac (macOS 26.5, Python 3.12):

• CLI, a new process per file: 14.5× to 1,660× faster, median 408×; 218 of the 231 files are at least 100× faster. The largest inputs are where the gap is smallest: the Llama 2 paper (26.6 s vs 1.6 s) and Tika's jxl.xls (18.2 s vs 1.1 s) are 16× faster. Python's start-up alone is 1.9 s; markitdown-rs answers --version in 4 ms.
• In process, warm, Python module against Python module: 4.7× (a large HTML page) to 48× (a small XLSX); 29× on the Llama 2 paper, 26× on jxl.xls.
• Peak memory is a quarter to a sixth of Python's: 79 MiB vs 440 MiB on the Llama 2 paper, 157 MiB vs 668 MiB on jxl.xls.

The tables below are from the repository's own benchmark corpus, generated by scripts/bench.sh. Python markitdown 0.1.9 (CPython 3.12.12, versions pinned in oracle/uv.lock) against markitdown-rs 0.1.0 (release build, rustc 1.99.1) on the same machine, run back to back on 2026-10-22: Apple M2 Max, 12 cores, macOS 26.5. The machine was shared with other jobs (load average 16–47), so treat wall times as indicative; CPU times track them closely. Raw output: bench/results-2026-10-22.txt.

CLI end to end, a new process per run (median of 10; Python on large.docx and report.pdf: 3):

In process, warm, excluding Python's 412 ms import (median of 5 Python and 10 Rust runs):

Peak RSS: large.docx Python 504,446,976 bytes, markitdown-rs 130,924,544; report.pdf Python 156,663,808, markitdown-rs 45,023,232. The markitdown binary is 4,524,288 bytes; the virtualenv with the markitdown extras used here is 328 MB.

All benchmark outputs were checked to be byte-identical before timing. To reproduce (needs uv and Rust):

scripts/bench.sh

It generates the corpus (scripts/gen_bench_corpus.py, fixed seed), builds the release binary and prints every table above.

Parity testing

• oracle/ is a uv project pinning markitdown 0.1.9 and its dependencies (mammoth 1.13.0, pdfminer.six, pdfplumber, pandas, openpyxl, xlrd, markdownify 1.2.3, python-pptx 1.0.2, ...).
• scripts/gen_snapshots.sh runs the Python CLI on every file in tests/fixtures/ and stores its stdout in tests/snapshots/ (or the error when Python fails, in which case markitdown-rs only has to not crash). scripts/gen_url_snapshots.py does the same for tests/url/cases.tsv, converting fixtures with the stream info of a URL response.
• cargo test converts every fixture and compares with the snapshot byte for byte. Accepted differences are listed with a reason in tests/known-diffs.txt; the test fails if a listed file starts matching.
• CI runs the tests on Linux, macOS and Windows and regenerates the snapshots from the pinned Python to prove they are current.

Fixtures come from the test corpora of markitdown, mammoth, python-pptx and Apache POI (real documents from bug reports), DP-Bench PDFs, Project Gutenberg EPUBs and Wikipedia pages, plus generated files that each target one construct (scripts/gen_*_fixtures.py); sources and licenses are in tests/fixtures/SOURCES.md.

How it works

Each converter ports the Python libraries markitdown calls, down to their string handling (Python's str.strip(), repr(float), urllib.parse, html.unescape), rather than wrapping or approximating them:

crates/markitdown-core      library
  convert.rs, detect.rs     markitdown's converter loop and content sniffing
  docx/                     mammoth's reader and HTML tree, OMML to LaTeX
  pptx/                     python-pptx's object model, charts
  pdf/                      pdfminer.six (parser, fonts, CMaps, layout) and pdfplumber words
  sheet/                    openpyxl/xlrd values, pandas frames and to_html
  html/, markdownify.rs     html.parser + BeautifulSoup tree, markdownify
  youtube.rs, epub.rs, msg.rs, rss.rs, zipfile.rs, plain.rs, media.rs
crates/markitdown-rs        the `markitdown` binary and `markitdown mcp`
crates/markitdown-py        the `markitdown` Python module (PyO3, maturin)

PDF data tables (glyph names, standard-14 metrics, CJK CMaps) are generated from pdfminer.six by scripts/gen_pdf_data.py; the CMaps are stored compressed (0.6 MB).

License

MIT. markitdown-rs ports behaviour from markitdown (MIT, Microsoft), mammoth (BSD-2-Clause), python-pptx (MIT), markdownify (MIT), pdfminer.six (MIT), pdfplumber (MIT), pandas (BSD-3-Clause), openpyxl (MIT), xlrd (BSD), BeautifulSoup (MIT) and CPython (PSF); the OMML conversion follows markitdown's adaptation of dwml, and the w:sym table is generated from mammoth's. The CMap data comes from pdfminer.six, which derives it from Adobe's CMap resources (BSD-3-Clause). Not affiliated with Microsoft.

• MCP Registry name: mcp-name: io.github.majiayu000/markitdown-rs