Picture in. PDF/A out. Capsule hashed. GPU-free. Command::Convert · SHA-256 ×3 · i32 1/7200" · PNG/JPEG → PDF/A XObject · not OCR. MinerU parses a scan. DocAgent embeds a PNG.

vs the field · Quickstart · What ships · API · Docs

Unique vs the locked ten — Command/Query/Event · capsule SHA-256 ×3 · prove same bytes. Not OCR. WeasyPrint tightness not claimed.

MinerU, Docling, Pandoc, Marker, Unstructured, PyMuPDF, WeasyPrint, ONLYOFFICE, python-docx, and LibreOffice do not ship an agent API, a three-hash capsule, and byte-identical replay together.

• Picture in, PDF/A out. PNG/JPEG on Markdown, HTML, DOCX, ODT paint as a PDF/A XObject. We embed the bytes. We do not OCR.
• Integer layout. Coordinates are i32 at 1/7200 inch. Same fonts → same bytes. GPU-free.
• Hangul is a regional codec on the same IR (images —), not the product name.
• Not claimed: OCR, VLM, OmniDocBench, WeasyPrint print-CSS tightness (table/paragraph still open).

Hashes below are the SHA-256 ×3 in letter-capsule.json.

_{examples/letter.md → PDF/A + HTML + capsule. Teal harbor mark under the H2. JSON b468b37f…; trio/capsule panel still b867db0a… (shots lag — not a fake hash). WeasyPrint tightness not claimed closed.}

Hash	SHA-256 (letter-capsule.json)
input	b0f4423728db4238779b3bf6cbd238d4d0d637f733c078322be6ff23398bbcdf
plan	4b7360eb1b7a9bec20cc35d5c79f2156c22d2712ab62ededfb60e328b8f16d3a
output	b468b37f1c3d8da7e4aac7e1a8417b46c29a57863817f9df16e8d4a6cc89f0f8

vs the locked ten

Locked against the ten most-starred document repositories an agent actually meets (star counts 2026-09-06). Agent-runtime matrix, not an OCR bake-off. No OmniDocBench numbers.

Tool	Job	C/Q/E	Capsule	Same bytes	i32 layout	GPU-free	Hangul
DocAgent	Agent document runtime	yes	SHA-256 ×3	yes	1/7200"	yes	yes
MinerU 79k	PDF/Office → LLM markdown
Docling 66k	Gen-AI parse + export					optional
Pandoc 46k	Universal markup convert					yes
Marker 40k	PDF → markdown/JSON
Unstructured 15k	Document ETL for LLMs
PyMuPDF 11k	PDF extract/convert					yes
WeasyPrint 10k	HTML/CSS → PDF					yes
ONLYOFFICE 6.9k	Collaborative office suite					yes
python-docx 5.7k	Word DOCX read/write					yes
LibreOffice 4.3k	Headless office convert					yes

North star: drop examples/letter.md (or DOCX/ODT) that contains a PNG or JPEG, and get PDF/A + HTML + Markdown + DOCX + ODT whose bytes match the next run, with three hashes in the capsule. The picture is a PDF/A XObject. Not OCR.

Images: yes PNG + JPEG on Markdown / HTML / PDF/A / DOCX / ODT. — on Hangul. The demo letter is ![Harbor mark](docs/assets/mark.png). First-screen convert shot above shows the teal mark under the H2. Demo-image gap closed. Print-CSS tightness vs WeasyPrint is not claimed closed.

Quickstart

git clone https://github.com/kevin9327/docagent
cd docagent
cargo test --workspace
cargo run -p docagent-cli -- convert examples/letter.md \
  --pdf out.pdf --html out.html --odt out.odt --docx out.docx --capsule cap.json
cargo run -p docagent-cli -- prove examples/letter.md

examples/letter.md contains ![Harbor mark](docs/assets/mark.png). Convert paints that teal mark under the H2 as a PDF/A XObject. prove converts twice (PDF/A + HTML) and exits 0 only when the bytes and the three SHA-256 hashes match. The prove JSON (letter-prove.json) is a separate artifact from the convert capsule in the table above — those plan/output hashes stay distinct (928b4745… / 2607bb6f…). Do not copy them onto letter-capsule.json.

Binaries: docagent (CLI) · docagentd (daemon) · MCP + WIT adapters over the same Engine.

use docagent_api::{Command, Engine, ExportTarget};

let mut engine = Engine::new();
let out = engine.execute(Command::Convert {
    input: std::fs::read("examples/letter.md")?,
    input_kind: None,
    targets: vec![ExportTarget::PdfA, ExportTarget::Html, ExportTarget::Markdown],
})?;
// out.capsule.{input_hash, plan_hash, output_hash}

---

Why this exists

MinerU, Docling, Marker, and Unstructured turn documents into tokens for models. Pandoc, python-docx, WeasyPrint, ONLYOFFICE, and LibreOffice turn documents into other documents. DocAgent is the runtime in the middle that an agent can call, replay, and prove.

What ships

Honest codec coverage — not a wishlist. Full table: docs/src/capabilities.md.

Mark	Markdown	HTML	PDF/A	DOCX	ODT	Hangul
Lists (ul/ol, nested)	yes	read ul/ol/tasks	markers	w:numPr	text:list	yes
Links	[text](url)	<a href>	URI /Link	w:hyperlink	text:a	yes
Strike	~~text~~	<s>	fill	w:strike	line-through	yes
Quotes	>	<blockquote>	quote bar	Quote	Quote	yes
Code	`span` / fence	<code> / <pre>	fills	Code	Tcode	yes
Thematic break	---	<hr/>	rule fill	HorizontalLine	HorizontalLine	yes
Tasks	- [x]	checkbox	task box	[x]	[x]	yes
Highlight	==mark==	<mark>	fill	w:highlight	Tmark	yes
Super / sub	^ ^ / ~ ~	<sup> / <sub>	baseline	w:vertAlign	Tsuper / Tsub	yes
Underline	__text__	<u>	fill	w:u	Tunder	yes
Images (PNG, JPEG)		<img>	XObject	a:blip	draw:image	—
Table header	GFM | --- |	<th> / <thead>	header fill	w:tblHeader	THcell	yes

Images are the first-class row. PNG and JPEG round-trip on Markdown, HTML, PDF/A, DOCX, and ODT. They stay — on Hangul. That dash is real.

HTML read recovers data-URI <img>, lists, <th>/<thead>, <blockquote>, <pre>, <a href>, and <mark> into the same IR. http(s): src is skipped. This is not a full HTML5 engine.

No spreadsheet. No slides. No cursor. No undo stack. No OCR. Flow documents only.

Convert shots

Files from examples/letter.md: PDF/A · HTML · DOCX · ODT · Markdown · capsule JSON.

![Harbor mark](docs/assets/mark.png)

PDF/A from the integer layout engine. Teal harbor mark under Delivery confirmation — order 48291. GPU-free. Same fonts → same bytes.

HTML from the same IR. Teal harbor mark under the H2 as <img>. read recovers data-URI images. Not HTML5. GPU-free.

_{Capsule from examples/letter.md: input / plan / output SHA-256. JSON b468b37f…; shot still shows b867db0a… (shots lag — not a fake hash). Two converts, same fonts, same bytes. GPU-free. Deterministic. Not OCR.}

The test is letter_md_is_byte_identical_across_two_runs. Hashes above are the hashes in letter-capsule.json.

Surface	What it is
IR	Format-blind document model. Layout never sees “docx” or “hwp”.
Layout	One engine. Coordinates are i32 at 1/7200 inch. f32 is forbidden.
PDF/A	Archival print from the integer plan. PNG and JPEG are XObjects. Two runs, same fonts, same bytes.
DOCX / ODT	Word and LibreOffice files. Headings, lists, links, tables, PNG + JPEG (a:blip / draw:image).
Markdown / HTML	Agent-native text and semantic HTML. . HTML <img> write / data-URI read. Not HTML5.
Capsule	Three SHA-256 hashes on every Command so a retry is evidence, not hope.
Hangul	Regional codecs on the same IR, not the product name. Images on Hangul are still —.

Command / Query / Event

Command  →  Convert | Import | Export | LayoutDocument
Query    →  EngineVersion | LastCapsule | LastDocument | CapsuleByInputHash
Event    →  CommandCompleted { capsule_id } | Warning

Every Command writes a capsule. CLI, daemon, MCP, and WIT are thin adapters over the same Engine. See AGENTS.md and docs/src/intro.md.

flowchart LR
  A[DOCX / ODT / MD / HTML + PNG/JPEG] --> B[docagent-model IR]
  B --> C[docagent-layout<br/>i32 @ 1/7200 in]
  C --> D[PDF/A XObject]
  C --> E[HTML]
  C --> F[DOCX / ODT / MD]
  B --> G[capsule.json<br/>input · plan · output]

Loading

Crate graph

Codecs (docx, odt, md, html, hwp5, hwpx, hwp3, hml) do not import each other and do not import layout. Layout, paint, raster, and PDF do not know the source format.

cargo test --workspace
cargo clippy --workspace --all-targets -- -D warnings
cargo run -p docagent-lint
cargo run -p docagent-conformance -- --require-win

Docs

• Intro
• What ships
• Format hops
• ADRs — crate boundaries, integer layout, single engine, capsules

License

MIT. See LICENSE.