Univer stopped calling itself an office SDK and started calling itself a harness
dream-num/univer has been around a while as an open-source SDK for embedding spreadsheets, docs, slides, boards and relational tables into your own product. The tagline now reads: the Office Harness for AI Agents. It is trending at 202 stars today on 15.3k total, Apache 2.0.
The repositioning is backed by three things, and the third is the one nobody else ships. Agents inspect and modify office content through structured APIs rather than pixels. Agents verify their own output through content inspection, screenshots and layout diagnostics. And agents work in isolated drafts β worktrees β that a human reviews before merging. A companion repo, univer-mcp, drives Univer Sheets over MCP in natural language.
The worktree idea is the interesting bet. Everyone building agent tooling for documents has run into the same wall: the agent's edit is either applied or not, and if it is wrong you are restoring from a version history that was never designed for a machine making forty changes a minute. Borrowing git's branching model for a spreadsheet means the review step is structural instead of a vibe check after the fact.
The headless Node runtime matters for the same reason. If the whole office suite runs without a browser, an agent can open a workbook, run a calculation chain, screenshot a range to check its own layout, and close it, all inside a server-side loop. A spreadsheet stops being a UI the agent has to poke at and becomes a computation environment it can call.
https://github.com/dream-num/univer
← Back to all articles
The repositioning is backed by three things, and the third is the one nobody else ships. Agents inspect and modify office content through structured APIs rather than pixels. Agents verify their own output through content inspection, screenshots and layout diagnostics. And agents work in isolated drafts β worktrees β that a human reviews before merging. A companion repo, univer-mcp, drives Univer Sheets over MCP in natural language.
The worktree idea is the interesting bet. Everyone building agent tooling for documents has run into the same wall: the agent's edit is either applied or not, and if it is wrong you are restoring from a version history that was never designed for a machine making forty changes a minute. Borrowing git's branching model for a spreadsheet means the review step is structural instead of a vibe check after the fact.
The headless Node runtime matters for the same reason. If the whole office suite runs without a browser, an agent can open a workbook, run a calculation chain, screenshot a range to check its own layout, and close it, all inside a server-side loop. A spreadsheet stops being a UI the agent has to poke at and becomes a computation environment it can call.
https://github.com/dream-num/univer
Comments