Formats

One read call handles every format below, detected from the file's bytes, not its name. Everything is parsed natively in Rust (zip/XML, calamine, html5ever): no Python, LibreOffice, or cloud service.

Input	What you get
PDF	Reading-order Markdown with headings, lists, tables, and page anchors
Word .docx	Headings, formatting, links, lists, tables, footnotes, equations, tracked changes and comments
PowerPoint .pptx	One section per slide, with notes and chart data
Excel .xlsx .xls .ods · CSV/TSV	One table per sheet
EPUB	One section per chapter
Embedded images	Figures and pictures inside PDF, DOCX, PPTX, EPUB saved as files and marked in place
HTML and URLs	The main article only
Markdown, text, JSON	Unchanged, with pagination
Images	Metadata, EXIF, and OCR text
Audio / video	Metadata, chapters, subtitles, and transcripts

PDF

Reading-order Markdown: headings, paragraphs, lists, tables, sub/superscripts, <!-- page N --> markers, and bookmarks as an outline. Running headers and page numbers are removed. Image-only pages use local doc-VLM OCR after explicit model setup, otherwise installed tesseract. Both read pages at 300 dpi. Doc-VLM uses layout boxes and reading order, with tables as Markdown and formulas as LaTeX; its model workers run one page at a time. Tesseract pages run in parallel and their words are laid out like a text page, so scans get paragraphs and tables too.

How it works: anymd reads glyph positions rather than text runs. Glyphs are grouped into lines by baseline, which tolerates super- and subscripts. Word spaces come from the gaps between glyphs, measured against the font size and adjusted for letter tracking. A column-aware XY cut finds gutters between running text. Tables come from drawn lines where a table has them (a missing line between two cells makes a merged cell) and from aligned columns of whitespace where it does not. Wrapped cell text stays in its cell, stacked header lines become one header, and a header over several columns is kept with each of them. Text a reader cannot see (invisible text, or text in the colour of the box behind it) is left out. Pages are processed in parallel and isolated from each other, so one malformed page never fails the whole document.

pages selects PDF pages. For images, geometry, or JSON structure, use inspect.

Word

Headings, bold/italic, links, nested lists, tables with merged cells, footnotes, and equations as LaTeX. Pictures are exported as image files.

Tracked changes and comments become CriticMarkup, in the body, tables, text boxes, and footnotes:

In Word	In the Markdown
Inserted text, or moved text at its new place	{++new++}{>>Ana Lima (2026-09-29T14:05:00Z)<<}
Deleted text, or moved text at its old place	{--old--}{>>Ana Lima (2026-09-29T14:05:00Z)<<}
Deleted text next to inserted text	{~~old~>new~~}{>>Ana Lima (2026-09-29T14:05:00Z)<<}
A comment on some text	{==text==}{>>Ana Lima (2026-09-30T08:15:00Z): comment<<}
A comment on a point	{>>Ana Lima (2026-09-30T08:15:00Z): comment<<}
An inserted or deleted paragraph break	{++ or {-- around the blank line between the paragraphs, then its author and date

Every tracked change is followed by who made it and when, as a comment, which is how CriticMarkup tracks several authors. A substitution made by two people names both, the deletion's author first. Neighbouring changes join into one only when the same person made them at the same time. Tracked changes inside a comment's own text keep their marks but not their author, since a comment cannot hold another comment. A change inside a link's text or a bold or italic run stays inside it ([the {++new ++}page](url)), and a footnote whose reference was inserted or deleted is marked the same way, label and all.

The author and date of changes and comments come from w:author and w:date. The date is copied exactly as stored, never converted; Word writes the author's local time there even though it ends in Z. Comments on the same text follow it in the order their anchors appear in the document, which is where Word puts a reply after the comment it answers. Formatting-only changes are not shown. In a document with tracked changes or comments, text that happens to contain a CriticMarkup delimiter is escaped with a backslash ({\++), so it reads the same but opens no span. In equations and image descriptions a space goes inside the delimiter instead (-- }), which LaTeX ignores. A document with no tracked changes and no comments is written exactly as before, with nothing escaped.

Accept or reject every change

revisions (CLI --revisions) picks how tracked changes come out:

revisions	What you get
markup (default)	The CriticMarkup above: every change, who made it and when, and the comments
accept	The text as Word shows it after Accept All, with no markup and no comments
reject	The text as Word shows it after Reject All, with no markup and no comments

accept and reject apply to the body, tables (inserted or deleted rows and cells are kept or dropped), text boxes and footnotes. A paragraph break that goes away joins its paragraph with the next one, which keeps the next one's style, as in Word. reject also puts back formatting that a tracked change replaced. A document with no tracked changes and no comments reads the same under all three.

PowerPoint

One section per slide in deck order: titles, bullets, tables, chart data, speaker notes, and pictures (exported as image files). pages selects slides.

Spreadsheets

.xlsx, .xls, .ods, CSV, and TSV. One Markdown table per sheet, dates as ISO strings, capped at 2,000 rows per sheet. pages selects sheets.

Spreadsheet formulas use the results saved by the spreadsheet application; anymd does not calculate formulas. A formula without a saved result appears as an empty cell. Recalculate and save the workbook in your spreadsheet application before converting it if results are missing.

CSV and TSV accept UTF-8 (with or without a BOM) and UTF-16LE or UTF-16BE with a BOM, with no encoding option needed. The CLI, MCP server, and browser playground decode them locally before parsing quoted cells. Malformed BOM-marked UTF-16 returns an error asking you to re-export the file, rather than replacing damaged text. Blank rows and empty edge columns are omitted; rows beyond the cap are counted below the table.

EPUB

One section per chapter in spine order, plus title and author, with images exported as files. pages selects chapters.

HTML and URLs

The main article only: navigation, cookie banners, and sidebars are dropped. Relative links are resolved, and code blocks keep their language. URL fetches are guarded; see Security.

Markdown, text, JSON

Returned unchanged, with pagination and the token budget.

Embedded images

An agent can open a standalone image file itself, but not one inside a container. read therefore exports the raster images embedded in PDF figures, DOCX and PPTX pictures and EPUB images (images: "refs", the default; CLI --images refs), and marks where each sits in reading order:

![Figure 1: Quarterly revenue by region](/home/me/.cache/anymd/images/9f2c1a7be03d55a4.png)
<!-- image: 240x150, page 1 -->

• Files. Each image is written once to <cache>/images/<first 16 hex of its SHA-256>.<ext>, where <cache> is $ANYMD_CACHE_DIR, else ~/.cache/anymd, ~/Library/Caches/anymd or %LOCALAPPDATA%\anymd\cache. The same picture is one file however often it is read. Files unused for 30 days are deleted (reading a document touches the files it uses), and the folder is kept under 2 GiB, oldest first; this runs at most once a day, at CLI and server start. Nothing is written beside the source document, and an image over 50 megapixels is refused.
• Captions. In a PDF, the nearest line directly below or above the image (within about half an inch) that starts with Figure, Fig., Table, 圖, 图 or 表 and a number. Otherwise the alt text (DOCX and PPTX descr, EPUB alt), otherwise image.
• Decoration is skipped. Images smaller than 48 x 48 px, PDF images under 2% of the page, and a picture that appears on three or more pages, slides or chapters (logos, running headers, ornaments).
• A scanned page is not a figure. A PDF page with no usable text layer and one image over at least 80% of it is the page itself: it gets no image ref and goes to OCR as before.
• Which images. PDF image XObjects (JPEG as is; 8-bit gray, RGB and palette rasters as PNG), DOCX word/media, PPTX slide pictures in slide order, and EPUB <img> files. The comment says page N for a PDF, slide N for a deck and chapter N for a book.
• Budget. A reference costs only its own two lines against max_tokens; the image itself is never counted.
• inspect with operation: "structure" also lists a local PDF's images as embeddedImages: page, bounding box in points (origin bottom left), size, caption and cached path.

Known gap in this version: figures drawn as vector graphics (charts, diagrams made of lines and text), inline images, CMYK, JPEG 2000, JBIG2 and CCITT images, and soft-mask transparency are not exported.

Images

Dimensions and EXIF (camera, date, GPS), plus local OCR. anymd setup ocr explicitly installs the doc-VLM engine and weights; automatic OCR otherwise uses installed tesseract without downloading anything. See local OCR.

Audio and video

Duration, streams, chapters, and embedded and sidecar subtitles (SRT/VTT), via ffprobe/ffmpeg. With transcript: true (CLI: --transcript), a local Qwen3-ASR transcript.

Transcripts

transcript: true (CLI: --transcript) uses bundled transcribe-cpp 0.2.4 and Qwen3-ASR-1.7B Q8_0 for every language, with automatic language detection. Whisper is no longer an engine option. Audio and documents stay on your machine.

Install ffmpeg to extract audio. With download_asr_model: true (CLI: --download-asr-model), anymd fetches Qwen3-ASR-1.7B-Q8_0.gguf (2.19 GB), checks its exact size and pinned SHA-256, and atomically saves it in $ANYMD_CACHE_DIR/models, or the platform cache (~/.cache/anymd/models, ~/Library/Caches/anymd/models, or %LOCALAPPDATA%\anymd\cache\models). Later requests verify and reuse it. ANYMD_ASR_MODEL can point at a preinstalled copy of the same pinned model; other models and corrupt cached files are rejected, not silently substituted. The pinned download revision and checksum are in asr.rs. To work offline, install that exact model first. transcript: true alone never downloads models; missing weights produce an installation hint. No weights are fetched for ordinary metadata or subtitle reads.

Long audio is decoded one 20-second chunk at a time, keeping audio memory bounded and all timestamps on the original source timeline. The model is loaded once per transcript. Segment timestamps mark chunk boundaries; they are not estimated word positions. Digitally silent chunks are skipped. An inference error or truncation is reported rather than presented as a complete transcript.

For word timestamps, an optional CrispASR 0.8.38crispasr binary on PATH (or ANYMD_ALIGNER_BIN) runs standalone alignment only with Qwen3-ForcedAligner-0.6B Q8_0. It does not run ASR or replace transcribe-cpp. The aligner weights (986 MB) are also revision/size/SHA-256 pinned and fetched on first supported use only when download_asr_model is true; ANYMD_ALIGNER_MODEL accepts a preinstalled pinned copy. Supported language codes are en, zh, yue, ja, ko, fr, de, it, pt, ru and es. Where that optional runtime is unavailable, the language is unsupported, or alignment fails validation, anymd keeps the transcript and labels its timestamps as segment, not word. Word timings must be monotonic, inside the chunk and account for the transcript text. Output explicitly names word, segment or mixed granularity and includes millisecond start/end times.

anymd talk.mp4 --download-asr-model  # explicitly allow the model download
anymd talk.mp4 --transcript          # cached/preinstalled models only

download_asr_model: true / --download-asr-model also implies transcript. Both CLI download flags are presence switches; inline values such as =false are rejected rather than granting permission. MCP download options are booleans. The old download_whisper_model / --download-whisper-model spelling is accepted only as a compatibility alias for Qwen; no Whisper engine or weights remain. Old ANYMD_WHISPER_* settings are not used.

Japanese trade-off: on our 200-utterance FLEURS sample, Qwen trails the whisper-turbo benchmark control (5.93 vs 4.80 raw CER; 5.56 vs 4.59 with symmetric numeral/kana normalisation). We use one speech model for all languages and accept this gap. See ASR benchmarks.

Optional tools

anymd never needs these, but uses them when they are on your PATH:

Tool	Adds
tesseract	OCR for images and scanned PDF pages
ffprobe	Audio/video metadata and chapters
ffmpeg	Embedded subtitles and transcript audio
crispasr (optional, alignment only)	Qwen3-ForcedAligner word timestamps (see Transcripts)

Check what anymd found:

$ anymd doctor
anymd 6.0.0 (native Rust)
  tesseract    found      OCR for images and scanned PDF pages
  ffprobe      found      audio/video metadata and chapters
  ffmpeg       found      embedded subtitles and transcript audio
Transcripts (--transcript):
  ASR runtime    transcribe-cpp 0.2.4 (CPU, bundled)
  ASR model      not installed: Qwen3-ASR-1.7B-Q8_0.gguf (2185 MB, SHA-256 pinned)

Typical installs: brew install tesseract ffmpeg on macOS, apt install tesseract-ocr ffmpeg on Debian/Ubuntu. For other OCR languages, install the tesseract language pack (for example tesseract-ocr-chi-tra).