AI coding assistants are good at text and code. Hand them a .docx, a .pdf or a .pptx, though, and they either flatten it into plain text (losing the formatting) or write a one-off python-docx script that breaks on the first complex style.
We build OmniDoc, a free desktop editor for documents, spreadsheets, slides, images and PDF on Windows, macOS and Linux. It ships with a local MCP server, OmniDocAI, so an agent can call document tools directly and produce real Office files. This post shows what that looks like. (Disclosure: we're the team behind it.)
Setup
The server speaks MCP over stdio and runs on your own machine. No API key is needed. The desktop app's "AI interface" page installs it in one click into clients such as Claude Desktop, Trae, Qoder and Kiro. For any other MCP client (Cursor, for example), add it by hand:
{
"mcpServers": {
"omnidoc": {
"command": "<install folder>/OmniDocAI.exe",
"args": ["--mcp"]
}
}
}
What an agent can do with it
-
Write once, get every format.
omnidoc_create_udocsaves an editable document plus an HTML preview next to it. Addexports: ["docx", "pdf"]and you also get Word and PDF in the same call. -
Edit a Word file and hand back Word.
omnidoc_edit_docxchanges text, paragraphs, styles, comments and table cells in place. Everything it does not touch stays as it was. No import/export round trip. -
Read documents properly.
omnidoc_read_documentreturns the full text of a PDF or Word file in reading order: headings, lists, tables as rows, headers and footers, footnotes. For Word it also returns comments and tracked changes. -
Translate a whole file and keep the layout.
omnidoc_translate_filekeeps tables, formulas and images in place. You can ask for side-by-side bilingual output or the translation only, pick a style (academic, legal, medical, technical…) and turn on context translation for consistent terms. This one needs a signed-in OmniDoc account and uses credits. -
Slides and sheets.
omnidoc_create_pptx/omnidoc_edit_pptx, thenomnidoc_render_pptxto render the slides so the agent can check its own layout.omnidoc_edit_sheetandomnidoc_render_sheetdo the same for spreadsheets. -
Show a PDF in an AI tool. Many AI clients can preview HTML but not PDF.
omnidoc_pdf_to_htmlconverts a PDF to a single HTML file offline in about a second. -
Share a link.
omnidoc_publish_filepublishes a document as a web page with an access code by default. - Work with Zotero. Search the library, add items with their PDFs, write notes and tags, and attach translated files to the original item.
A typical request
"Read report.pdf, write a two-page summary, and give me Word and PDF."
The agent calls omnidoc_read_document, then omnidoc_create_udoc with exports: ["docx", "pdf"]. You end up with three files side by side: the editable source, the Word file and the PDF.
Limits, honestly
- Scanned PDFs need text recognition first for stable results.
- Translation and web publishing need an account. Local editing and format conversion are free.
- Check numbers and citations in anything an AI writes.
Try it
Download from omnidoc.top (Windows, macOS, Linux), open the "AI interface" page and connect your MCP client. Questions are welcome in the comments.
Top comments (0)