OmniDoc: document translation and a free AI office suite

Translate PDF, Word, PowerPoint, LaTeX and 11 more formats into 38 languages with the layout intact. Four free offline editors, and AI assistants that make the files for you.

38 languages · 15 file formats · Offline editing · MCP for AI assistants · API · Windows · macOS · Linux · Web

• Website: https://omnidoc.top/
• Download the free desktop app: https://download.omnidoc.top/
• API documentation: https://docs.omnidoc.top/

This package is the OmniDoc Python SDK. It calls the OmniDoc API, so it needs an OmniDoc account and an API key (see Quick start).

Why OmniDoc

The world's most powerful AI document engine, all in one app.

The first product to put PDF layout reconstruction, interactive dynamic documents and AI-native office work into one app: translated PDFs keep their layout, documents become web pages, and AI builds your Word, Excel and PowerPoint files. The desktop app is free.

• Papers and contracts translated, layout intact. Tables, formulas and footnotes stay in place, scans and LaTeX papers work too, and bilingual output is ready to hand in.
• The world's most powerful PDF layout reconstruction. Keep every page 1:1 as it looks, or rebuild paragraphs, headings, lists and reading order into a Word file you can edit right away. Offline on your computer.
• Documents that move, shared as web pages. Put charts, animations and flowcharts in the document and readers can interact with them; export to Word or PDF whenever you need.
• One sentence, and AI makes the Word, Excel or PPT. Connect Claude, Cursor or another AI assistant to write documents, build spreadsheets and decks, and convert PDFs on your own computer, at no cost.

Three core technologies behind every OmniDoc product

Structure extraction that keeps the layout, documents that are interactive web pages, and an office suite that AI assistants operate directly.

Production-grade structure extraction that keeps the layout. OmniDoc reads PDFs, Word files and scans into their real structure: headings, paragraphs, lists, tables, formulas, headers, footers, footnotes and reading order. Translation and conversion put every element back where it was, so a translated paper keeps its equations and a converted PDF keeps its pages. PDF to a web page with selectable text in about one second on your own computer; scanned and complex pages go through visual structure extraction.

From a fixed PDF layout to a flowing document. A PDF only records where each piece of text sits on the page; it has no paragraphs or headings. OmniDoc rebuilds paragraphs, headings, lists, tables and reading order from those scattered blocks and writes an editable, flowing document, or keeps the original layout page by page when you need it.

UniDoc: documents that are interactive web pages. A UniDoc document can hold interactive charts, animations, flowcharts, tabs, cards and numbered equations that readers click, drag and play. Publish it and it stays interactive at a public web address; when you need a static copy, export the same file to Word, PDF or lossless HTML. AI assistants can write a whole interactive document in one go. Native .udoc format, about 9 ms from keystroke to layout, and the same editor in the desktop app and the browser.

An AI-native office suite: your assistant does the work. OmniDoc Desktop ships with an MCP server of more than 40 tools. Claude, Cursor and other MCP apps use it to create and edit Word, Excel and PowerPoint files, convert PDFs, translate, publish and manage Zotero papers, and to check their own work: every edit returns layout problems, recalculated formulas or slide previews. Runs on your computer: editing and conversion are offline and free; cloud services such as translation use credits.

Ask your AI assistant, get a finished file

Install OmniDoc Desktop and add it to Claude Desktop or Cursor in one click, or to any app that supports MCP. Your assistant then works on real files on your computer.

• Word documents. Write reports, papers and résumés from templates, import an existing .docx, and save Word, PDF and lossless HTML in one call. Figures, tables and equations are numbered, and layout problems are reported back.
• Excel workbooks. Build or change .xlsx files with formulas that are recalculated and checked, native Excel charts, pivot tables, conditional formatting, filters and sparklines.
• PowerPoint decks. Create decks from your own template or a blank 16:9 deck, fill layouts with text, tables, charts and pictures, add animations and transitions, and catch overflowing text on every slide.
• PDF conversion. Convert PDFs offline to Word with the page layout kept or as editable paragraphs, to HTML with selectable text, or to images; merge, split and compress PDFs.
• Read, translate and publish. Read Word and PDF files in reading order, translate files with the same engines as omnidoc.top, and turn any document into a public web link.
• Papers and Zotero. Search your Zotero library, add papers with their PDFs, notes and tags, and fetch a paper's full text from open-access sources or your own institutional sign-in.

Editing and conversion run offline at no cost. Translation and other cloud services use credits, and your assistant acts as the account signed in to OmniDoc, so there is no API key to paste.

APIs for developers

• Document API. REST endpoints to upload files, run translation jobs, track them and download results, plus file tools, glossaries, instant text translation and audio and video jobs.
• OpenAI-compatible AI API. Point any OpenAI SDK at https://omnidoc.top/v1 and call Claude Haiku 5.5, with streaming. Each call is charged its actual model cost: 1 yuan of cost is 1,000 credits.
• MCP server. The local MCP server in OmniDoc Desktop gives AI agents more than 40 document tools that work offline, with no API key.

Pricing

Translation is charged for what is actually delivered, roughly 4 credits per page. The smallest pack is 3,000 credits for ¥9.9 (about 750 pages), and bought credits never expire. The editors and offline conversion are free.

---

Python SDK

Install

pip install omnidoc

Requires Python 3.9 or newer. The only dependency is requests.

Quick start

Create an API key (it starts with sk_) in the API center at https://omnidoc.top/app/api-keys and put it in the environment:

export OMNIDOC_API_KEY="sk_your_key"

from omnidoc import OmniDocClient

client = OmniDocClient()  # reads OMNIDOC_API_KEY
files = client.translate_file("report.docx", "en", output_dir="out")
print(files)

translate_file uploads the file (presign, direct upload, finalize), creates a translation job, polls until it ends and downloads every artifact. A failed job raises JobFailedError; quota_exceeded is true when the balance was too low to start, in which case nothing is charged.

Step by step

doc_id = client.upload("report.docx")
job = client.create_job(doc_id, "en", output_mode="bilingual")
job = client.wait_job(job["id"], poll_interval=3, timeout=1800)
for artifact in client.list_artifacts(job["id"]):
    client.download_artifact(artifact["id"], "out")

Each create_job call generates a new Idempotency-Key; internal retries reuse it, so a job is never created or charged twice. To retry the same submission yourself, pass the same idempotency_key.

Account and other endpoints

client.get_job(job_id)           # job status (free)
client.cancel_job(job_id)        # cancel a job
client.credit_balance()          # credit balance (free)
client.credit_operations()       # credit history
client.usage_summary()           # usage summary

# Endpoints not wrapped yet: call them by the path shown on docs.omnidoc.top
client.get_json("<path>", params={...})
client.post_json("<path>", json={...})

Errors and retries

• API errors raise OmniDocError with status_code, code, message and request_id (the X-Request-Id response header).
• 429 and 5xx responses are retried using Retry-After or exponential backoff, up to max_retries times (default 4).
• AUTH_API_KEY_QUOTA_EXCEEDED is not retried: the key's budget is used up and has to be raised first.

Webhook signatures

from omnidoc import parse_webhook

event = parse_webhook(raw_body, request.headers["X-Webhook-Signature"], secret)

Verify with the raw request body bytes. A bad signature or an old timestamp raises ValueError. De-duplicate events by data.job_id and data.status.

Security

• Keep the key on your server or in a controlled environment, never in source code, logs or a public repository.
• The key is sent only to the API address, never to the direct-upload upload_url.

Development

pip install -e ".[test]"
pytest

The tests use a fake HTTP session: no API key and no network.