Comparison · Same 38 public documents · Open benchmark

anymd vs Docling

Both tools turn documents into Markdown for LLMs, and both are free and MIT licensed. Docling is a Python library and CLI, started at IBM Research, with layout and table models. anymd is one native Rust binary that is also an MCP server. On the 38 public documents of AgentDocBench, anymd scores 96.4 to Docling's 93.0, converts all 38 in 11.5 s instead of 2,432.4 s on the same 4-CPU machine, and returns 14% fewer output tokens.

Install anymd free    See the full benchmark

npx -y @sylphx/anymd setup --client cursor,claude-code,codex

At a glance

anymd Docling
What it is One native Rust binary: CLI and MCP server Python library and CLI, with a separate docling-mcp server
Install npx -y @sylphx/anymd setup, Homebrew, Scoop, Cargo or Docker pip install docling, Python 3.10 or newer; the standard install brings PyTorch and its layout and table models
MCP tools Built in: read, outline, search, inspect docling-mcp, installed separately
AgentDocBench overall 96.4 93.0
Time, all 38 documents, 4 CPUs 11.5 s 2,432.4 s
Output tokens, all 38 documents 368,659 430,148
Scanned PDFs Local OCR with tesseract, or a local document model after anymd setup ocr Several OCR engines, and vision-language models such as GraniteDocling
Audio and video Metadata, chapters, subtitles, and local speech-to-text after a one-time model download Speech-to-text transcripts; video with a transcript and representative keyframes
Output Markdown with <!-- page N --> anchors; JSON with geometry through inspect Markdown, HTML, lossless JSON (DoclingDocument), DocTags and more
Python API Prepared, not on PyPI yet Yes, with LangChain, LlamaIndex, Crew AI and Haystack integrations
Inputs only Docling reads — Email (EML, MSG), calendar files, Apple Pages, Numbers and Keynote, XBRL, LaTeX, OpenDocument text and slides
Licence MIT MIT code; each model has its own licence
Price Free, MIT. Optional anymd Pro, one payment Free

Measured on the same 38 documents

AgentDocBench checks what an agent needs from a converter: the words verbatim, in reading order, with tables intact, fast, and in as few tokens as possible. Scores are 0 to 100; higher is better.

AgentDocBench anymd 8.2.0 Docling 2.130.0
Overall 96.4 93.0
Documents converted 38/38 38/38
Sentences kept verbatim 97.7 94.4
Text F1 97.2 94.5
Reading order 98.8 94.4
Table cells F1 92.2 89.9
Time, all converted documents 11.5 s 2,432.4 s
Output tokens, all converted documents 368,659 430,148

Document by document, anymd scores higher on 22, Docling on 10, and 6 are tied.

By category

Category Documents anymd Docling
Papers (math) 3 96.7 91.5
Two-column papers 3 98.1 93.1
Reports 2 96.9 97.7
Financial tables 4 95.2 92.4
Forms 3 98.0 95.6
Scanned (OCR) 3 77.9 78.4
CJK 3 99.1 97.4
Slides 4 96.0 93.8
Spreadsheets 4 100.0 85.8
Word 3 99.9 99.9
EPUB 3 98.9 99.7
HTML 3 99.6 90.5

Where Docling scores higher: Reports (97.7 vs 96.9); Scanned (OCR) (78.4 vs 77.9); EPUB (99.7 vs 98.9); paper-evsched (98.9 vs 98.4); report-crs-dod-approps (96.1 vs 94.8); report-plz (99.3 vs 99.0); invoice-borderless (99.5 vs 94.8); irs-w9 (99.8 vs 98.9); scan-nasa-budget (77.2 vs 71.6); slides-epa-1bp (92.1 vs 91.7); slides-markitdown-pptx (99.0 vs 98.5); slides-nasa-clarreo (95.2 vs 94.5); epub-markitdown (99.3 vs 96.8).

Docling is the closest tool on accuracy. The gap on this benchmark is 3.4 points overall; the bigger difference is time. Docling runs a layout model on every page and a table model on every table, so on a CPU each document took about a minute on average, while anymd reads the PDF's own text positions.

How it was measured

Both tools ran on GitHub-hosted Linux x86_64 runners with 4 CPUs: Docling 2.130.0 through its CLI with its default pipeline (layout and table models, OCR on bitmap regions, CPU only) on 2026-09-25, and anymd 8.2.0 on 2026-10-01. Time is the wall time of a fresh process per file, start-up and model loading included: one timed run after a warm-up for Docling, and the median of three runs for anymd. A GPU, a long-running Docling process or its vision-language pipeline would change Docling's time and scores; none of them was measured, and neither was Docling 2.150.1, released after the run. Tokens are counted with o200k_base.

The benchmark lives in the anymd repository. The scoring does not know which tool produced an output, and the corpus, ground truth, scoring and raw results are public in bench/. Anyone can rerun it or submit a tool.

Why agent builders pick anymd

One binary, no Python, PyTorch or GPU

npx -y @sylphx/anymd setup adds anymd to Claude Code, Codex, Cursor, VS Code, Claude Desktop, Windsurf and Gemini CLI at once. Homebrew (brew install sylphxai/tap/anymd), Scoop (scoop install anymd from the SylphxAI/scoop-bucket bucket), Cargo (cargo install anymd) and Docker (ghcr.io/sylphxai/anymd) install the same binary. There is no virtual environment, no PyTorch install and no model download to convert a born-digital document.

Fast enough to call while the agent waits

On the same 4-CPU runner, anymd converted the 38 documents in 11.5 s and Docling in 2,432.4 s. An agent can convert a document in the same turn it asks for it, instead of handing a folder to a batch job.

More of the text, in order, in fewer tokens

anymd keeps more sentences verbatim (97.7 vs 94.4) and in reading order (98.8 vs 94.4), and returns 368,659 tokens to Docling's 430,148. The largest category gaps are spreadsheets (100.0 vs 85.8), HTML (99.6 vs 90.5), math papers (96.7 vs 91.5) and two-column papers (98.1 vs 93.1).

Built for agents

The MCP tools add what an agent needs for long files: <!-- page N --> anchors to cite, outline to jump to a section, read with a token budget and a cursor, and search across a folder. Files stay on your machine, and --allow-dir keeps the MCP server inside one folder.

When Docling is the better fit

  • You want an in-process Python library today, with its LangChain, LlamaIndex, Crew AI or Haystack integrations. anymd's Python package is prepared but not yet on PyPI.
  • You need a lossless structured document model (DoclingDocument JSON) or DocTags rather than Markdown.
  • You read email, calendar, Apple iWork, XBRL or LaTeX files, or want to run vision-language models such as GraniteDocling on a GPU.
  • Your corpus is mostly scans, designed reports or EPUB, where Docling scored slightly higher on this benchmark.

Switch in one line

docling report.pdf --to md --output out/   # Docling
anymd report.pdf -o report.md              # anymd, after npm install -g @sylphx/anymd

For MCP, replace the docling-mcp entry in your client with npx -y @sylphx/anymd, or run npx -y @sylphx/anymd setup to add it to every client. With Docker:

docker run --rm -v "$PWD:/data" ghcr.io/sylphxai/anymd report.pdf > report.md

More install options are in Getting started.

Check what your agent cites with anymd Pro

Converting a file is free. When an agent quotes a PDF or a video, anymd Pro takes you back to the source:

  • Cite-check compares up to 100 quotes per request with the page and area each one cites, and reports whether it matches, doesn't match, or needs more evidence. It uses the text and its position on the page, not an AI opinion.
  • Video timelines and frames show scenes, chapters and subtitle or transcript timings, and the frame behind a cited moment with its actual decoded timestamp and a SHA-256 hash of the image.

One-time US$29 purchase. No subscription. Sign in, pay securely, then run anymd login with the same account to activate Pro on your machine. After activation, Pro works offline.

See Anymd Pro

FAQ

Is anymd a drop-in replacement for Docling? For Markdown output, mostly: anymd report.pdf -o report.md replaces docling report.pdf --to md, and npx -y @sylphx/anymd setup adds anymd to your MCP clients. Code that imports docling cannot switch today, because anymd's Python package is not on PyPI yet and anymd does not produce DoclingDocument JSON.

Is anymd faster than Docling? On AgentDocBench, anymd converted all 38 documents in 11.5 seconds and Docling in 2,432.4 seconds on the same 4-CPU runner, each timed as a fresh process per file with start-up and model loading included. Docling on a GPU was not measured.

Does anymd need Python, PyTorch or a GPU? No. anymd is a native Rust binary. The npx install needs Node 18 or newer; Homebrew, Scoop and Docker install the same binary without Node.

Which one is more accurate? They are close: 96.4 overall for anymd and 93.0 for Docling. anymd scores higher on 22 documents and Docling on 10. Docling is ahead on scanned documents, designed reports and EPUB.

Is the benchmark biased toward anymd? It lives in the anymd repository, so you can check it: the scoring does not know which tool produced an output, the corpus and ground truth are public, and the categories where anymd loses are published. Anyone can rerun it or add a tool.

Is anymd free? Yes. anymd is MIT licensed, including the CLI, the MCP server, OCR and transcripts. anymd Pro is an optional one-time purchase, US$29 or the local price in 16 currencies, that adds PDF cite-check and video timelines and frames.

Docling is an LF AI & Data Foundation project started at IBM Research; anymd is not affiliated with or endorsed by IBM or the Docling project. Benchmark figures come from the AgentDocBench results committed in the anymd repository. Docling details come from its README.