anydoc-serve

Office documents, PDFs and scans to Markdown over HTTP and MCP, in one Docker container, with local OCR for English, Japanese and Thai.

anydoc-serve wraps firecrawl/anydoc, the fast Rust converter for Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV and PDF, in a server you can run anywhere:

docker run -p 8080:8080 ghcr.io/bon5co/anydoc-serve
curl -F file=@report.docx http://localhost:8080/v1/convert

What it adds over plain anydoc

OCR is Tesseract 5 reading every configured language in one pass, so a page that mixes Japanese, Thai and English is read correctly line by line. It was picked by measuring it against RapidOCR/PaddleOCR on English, Japanese, Thai and mixed-script pages: see the benchmark.

Related MCP server: Markitdown MCP Server

Quickstart

docker run -d -p 8080:8080 -e API_KEY=change-me ghcr.io/bon5co/anydoc-serve

or docker compose up -d with the docker-compose.yml in this repo.

Convert an upload:

curl -H "Authorization: Bearer change-me" -F file=@scan.pdf http://localhost:8080/v1/convert

{
  "markdown": "# Quarterly report\n\n...",
  "metadata": {
    "filename": "scan.pdf", "bytes": 460689, "format": "pdf", "page_count": 4,
    "ocr_pages": [3, 4], "ocr_engine": "tesseract", "ocr_langs": ["eng", "jpn", "tha"], "duration_ms": 1735.0
  }
}

ocr_pages lists the 1-indexed pages that went through OCR; every other page came from anydoc.

Just the Markdown:

curl -H "Authorization: Bearer change-me" -H "Accept: text/markdown" -F file=@report.docx http://localhost:8080/v1/convert

Have the server fetch a URL:

curl -H "Authorization: Bearer change-me" -H "Content-Type: application/json" \
     -d '{"url": "https://example.com/report.pdf"}' http://localhost:8080/v1/convert

Interactive docs are at /docs (OpenAPI JSON at /openapi.json).

API

POST /v1/convert

• multipart/form-data: file (required), format, ocr, output

• application/json: {"url": "...", "format": ..., "ocr": ..., "output": ...}

Errors answer {"error": {"code": ..., "message": ...}}: 400 bad request or unfetchable URL, 401 missing or wrong key, 413 over MAX_UPLOAD_MB or OCR_MAX_PAGES, 415 unknown format, 422 encrypted, malformed, or needs OCR with OCR off, 504 over CONVERT_TIMEOUT_S.

GET /health answers {"status": "ok", "ocr": {...}, "auth": true} and never needs the key.

MCP

The server speaks MCP over streamable HTTP at /mcp, with one tool, convert_document(url | content_base64, filename?, format?, ocr?), which returns the Markdown.

Claude Code:

claude mcp add --transport http anydoc http://localhost:8080/mcp --header "Authorization: Bearer change-me"

Clients configured with an mcpServers JSON block that supports remote HTTP servers (Cursor, for example):

{
  "mcpServers": {
    "anydoc": {
      "type": "http",
      "url": "http://localhost:8080/mcp",
      "headers": { "Authorization": "Bearer change-me" }
    }
  }
}

Configuration

Resource use

Measured on linux/amd64, 2026-09-28:

• Image: 376 MB on disk (94 MB compressed download). Tesseract models for English, Japanese, Thai and script detection are 18 MB of that.

• Memory: about 305 MiB RSS after a round of conversions (docker stats); each conversion slot loads about 90 MiB of OCR models on first use.

• Speed: a born-digital .docx converts in about 2 ms; OCR takes about 1.1 s per page (median, 4 cores of an Intel i5-10500T). A 4-page PDF with 2 scanned pages took 1.5 to 1.7 s.

OCR benchmark

Character error rate on synthetic 200 dpi pages: degraded scans, two pages per language plus one page each mixing Japanese+English, Thai+English and all three. Median seconds per page. Full method, clean-render numbers, every engine row and reproduction steps: docs/ocr-benchmark.md.

RapidOCR reads pure Japanese far better and two-script pages slightly better. It fails the three-script page, drops lines of English, and is 3 times slower than the server as shipped. Synthetic pages give exact ground truth but are not your documents: treat these as a comparison between engines, not an accuracy promise.

Deploy on Railway

One service, no volume. API_KEY is generated at deploy (read it from the service's Variables tab) and OCR_LANGS defaults to eng,jpn,tha. Measured on Railway: 246 to 263 MiB of RAM idle, 272 MiB after a round of conversions.

Development

uv sync
uv run pytest                      # unit tests, no Tesseract needed
docker build -t anydoc-serve .
docker run -d -p 8080:8080 -e API_KEY=test-key anydoc-serve
ANYDOC_SERVE_URL=http://127.0.0.1:8080 ANYDOC_SERVE_API_KEY=test-key uv run pytest -m integration

Test fixtures anydoc-text.docx and anydoc-text.pdf are copied from anydoc's own test corpus (MIT); the scanned fixtures are generated by tests/fixtures/make_fixtures.py.

Credits and license

The document conversion is all anydoc by Firecrawl (MIT). OCR is Tesseract (Apache 2.0) through tesserocr (MIT); PDF pages are rendered with pypdfium2. anydoc-serve is not affiliated with Firecrawl.

MIT, see LICENSE.