An MCP server that hands AI agents actual .docx and .xlsx files — headings, tables, frozen headers, and right-to-left text that actually works.
Copy it now — it is not shown again.
A language model writes words. It cannot assemble a file. Under the hood a Word document is a zip of XML parts where every byte has to land in the right place — so the model returns formatted text and hopes somebody pastes it into Word.
ההכנסות עלו ב־12% לעומת הרבעון הקודם, בעיקר בזכות אזור הצפון.
| אזור | הכנסה |
|---|---|
| צפון | 120 |
| דרום | 90 |
Add it to any MCP client — Claude, Cursor, your own agent. Paste your key and it is done.
// claude_desktop_config.json / .cursor/mcp.json { "mcpServers": { "docmcp": { "type": "http", "url": "/mcp", "headers": { "Authorization": "Bearer YOUR_KEY" } } } }
Then the agent just calls it:
create_docx({ title: "דוח רבעוני", rtl: true, blocks: [ { type: "heading", text: "סיכום", level: 2 }, { type: "paragraph", text: "ההכנסות עלו ב־12%" }, { type: "table", columns: ["אזור", "הכנסה"], rows: [["צפון", "120"], ["דרום", "90"]] } ] }) → https://…/f/report.docx
It does one job properly instead of ten badly.
Headings, paragraphs, bullet and numbered lists, tables, page breaks. Set rtl for Hebrew or Arabic.
Multi-sheet, bold frozen header row, auto-filter, sized columns. Numbers stay numbers.
Documents used and remaining this month. No surprise bills.
No seats. Agents are not people, and charging per seat stopped making sense.