The Spreadsheet Becomes the Agent's Workbench

Univer, the open-source Office SDK from the original Luckysheet team, has repositioned itself as the place where AI agents and humans edit the same files — and review each other's work like colleagues.
Every few months, the open-source world rediscovers that spreadsheets are hard. Not hard to render — any table library can draw a grid — but hard to be: a formula engine with dependency graphs and circular reference detection, number formatting that respects locales, undo stacks, collaborative editing, and the thousand small behaviors Excel users have muscle memory for. The graveyard of “open-source Excel killers” is well populated. Univer, the SDK from DreamNum, is the latest serious attempt, and its recent traction comes from a bet that is more interesting than another spreadsheet: that the next users of office software are not humans at all.

The project now bills itself as “The Office Harness for AI Agents,” and that phrasing is doing a lot of work. Univer is not a hosted app you log into. It is an SDK — a set of building blocks for embedding spreadsheet, document, and presentation editing into your own product, with a plugin architecture, a Canvas-based rendering engine, a formula engine, and a Facade API that runs identically in the browser and headless on Node.js. The same architecture that draws an editable grid on a web page can recalculate a workbook on a server with no UI attached. That isomorphism is the technical foundation for the agent story, and it is worth unpacking before the hype.
The second act of the Luckysheet team
Univer’s credibility does not come from nowhere. The team behind it built Luckysheet, the widely used open-source JavaScript spreadsheet, and the project site claims 44.5k GitHub stars across Univer and Luckysheet combined, plus “10+ yrs” building office tools. Luckysheet was a genuine phenomenon in the Chinese and international front-end communities — a single-package, drop-in spreadsheet that worked when nothing else did. But it had limits: a monolithic design, a rendering approach that strained on large workbooks, and an architecture that was difficult to extend into documents and slides.
Univer is the rewrite that answers those regrets. Where Luckysheet was one big package, Univer is a monorepo of composable plugins: a core SDK, a rendering engine, a formula engine, and separate packages for sheets, docs, slides, and their UI layers. Every capability — number formatting, filtering, data validation, comments — arrives as a plugin that can be added, removed, replaced, or lazy-loaded. The README is explicit that Univer is “not a spreadsheet file viewer only” but “a framework for building your own productivity surface.” That distinction matters: the customer here is a product team embedding office capability, not an end user looking for a free Excel.
The project also claims the top spot on the SpreadsheetBench leaderboard as of December 2025, per its own site. Self-reported benchmark leadership deserves the usual skepticism — the methodology is not explained on the landing page — but formula-engine correctness is exactly the kind of thing benchmarks can measure, and Univer clearly considers it a competitive front.
Why isomorphic is the whole point
The boring part of Univer is where the value lives. Most office SDKs in the market force an architectural choice: either a client-side editor (with all the fidelity and performance problems that entails) or a server-side document engine that converts and processes files away from the user. Apryse, a commercial competitor, markets exactly this split — client-side rendering and editing across web, mobile, and desktop, with no reliance on server-side conversion or LibreOffice. Thinkfree, another commercial player, structures its SDK around where the document engine lives: their infrastructure, your private cloud, or your air-gapped rack.
Univer’s answer is stranger and more ambitious: the same code, both places. The rendering engine draws to Canvas in the browser; the model, command, and formula layers run anywhere JavaScript does, including Node.js 18.17 and later. A scheduled job can fill a template, calculate, and export a report with no editor open — the site calls this “back-office automation” — while a human user sees the identical document model in an interactive view. There is no conversion step between the agent’s view of a workbook and the user’s, because there is only one model.
This is what makes the agent pitch more than marketing. The current wave of AI spreadsheet tooling mostly attacks the problem from the outside. Autosheet, the artist formerly known as GPT for Work, is an add-in for Excel and Google Sheets that specializes in bulk row processing — a thousand rows a minute, per its own marketing — and now has an agent that checks its own charts via screenshots before answering. A recent survey of agent-accessible spreadsheet tools frames the evaluation criteria well: API or MCP access so agents can read and write cells programmatically, bulk processing capacity, and auditability so a human can trace what changed. Quadratic, the MCP-native spreadsheet with Python and SQL in every cell, scores highly on all three by building the agent interface into the product from day one.
Univer is playing the same game from the infrastructure side. Rather than bolting an agent onto Excel, it offers the substrate an agent can inhabit: structured programmatic editing, output verification through content inspection and rendered screenshots, and — the most distinctive piece — worktree collaboration.
Pull requests for spreadsheets
The worktree idea is the part of Univer’s story that feels genuinely new. Borrowed directly from version control, it lets an agent work in an isolated draft of a document while the main version stays untouched. A human then reviews the agent’s changes — the marketing materials show diffs of a requirements doc, an event catalog spreadsheet, and a launch brief, side by side, with additions and deletions highlighted — and decides what to merge. The framing on the site is explicit: “Manage Office files and Web Apps the way GitHub manages code.”
This is a response to a real and unresolved problem in agentic work. When an AI agent edits a fifty-tab financial model, “trust me” is not an acceptable review process. The fast.io survey lists auditability as a first-class criterion precisely because most tools are bad at it — a chat log is not a diff. Univer’s bet is that the review primitive agents need already exists, and it is the pull request. Whether office documents diff as cleanly as source code is an open question — a changed formula cascading through a workbook is not a one-line change — but the direction is right, and the ecosystem around the repo shows it is being taken seriously: first-party integrations exist for DeepSeek Harness, WorkBuddy, and OpenClaw, a CLI for local agent workspaces, and a separate MCP project for driving Univer Sheets with natural language.
There is also a skills repository — reusable instructions for AI agents working with the SDK — which is a quietly telling artifact. When your documentation includes a corpus written for the agents that will integrate you, you have committed to a world where LLMs are a distribution channel.
The open-core tightrope
Univer is Apache-2.0, and the README is unusually candid about where the open source ends. The OSS repository covers the core SDK, the plugin system, the rendering and formula engines, the Facade API, and a substantial sheets surface: formulas, number formatting, filtering, sorting, data validation, conditional formatting, comments, find and replace. The commercial Univer Pro layer holds real-time collaboration, import/export, printing, charts, pivot tables, and server-side calculation.
This is a defensible boundary — collaboration servers are expensive to build and support — but it is also a boundary drawn right through the middle of the agent story. The README notes that live editing, shared revisions, and worktree workflows “require the corresponding Web SDK and collaboration capabilities; package availability and licensing vary by feature.” Read carefully: the headline agent workflow, the one the landing page leads with, sits partly on the paid side. The team deserves credit for stating this plainly and for publishing boundary principles — OSS bugs get fixed in the OSS repo, OSS docs must not imply Pro features are free — but a team evaluating Univer for an agent product should assume the compelling demo and the Apache-2.0 repository are not the same thing.
Maturity is uneven in ways the README admits. Sheets are “the most mature product surface today”; Docs and Slides share the architecture but “continue to evolve”; Slides packages are under active development; Bases — the structured-data product — is mostly an extensibility story in OSS with the actual database model living in Pro. PDF support is listed as “coming soon.” The compatibility targets are modern but not aggressive: Chrome 88, Firefox 90, Safari 14.1, with a required Intl.Segmenter polyfill for older runtimes. None of this is damning for an SDK at this stage, but it is a reminder that the breadth of the vision outruns the depth of every surface except sheets.
The landscape it lands in
Univer is not competing with Excel or Google Sheets; it is competing with the commercial office SDKs — Apryse, Thinkfree, WPS’s offering (a Reddit thread on embedding WPS documents in React suggests developers are actively shopping this category, though the thread itself was inaccessible) — and with the agent-native newcomers like Quadratic. Against the commercial SDKs, its differentiators are price on the OSS core, hackability, and the Node.js headless runtime. Against the agent-native tools, its differentiator is that it is a full office suite architecture rather than a spreadsheet with hooks.
The customer logos on the site — Genspark, Brightwave, Energent.ai, XpertAI, Hengchang — skew toward AI companies and BI-adjacent products, which fits: the buyers who need an embeddable office surface and an agent harness are the ones building agent products themselves.
One hazard of writing about Univer is the name. “Univer Video” is an unrelated Brazilian streaming service for Christian films and telenovas, available on Google Play with over a million downloads and a devoted audience requesting more English audio for its Paul the Apostle docudrama. The office SDK has no opinion on any of this, and neither, regrettably, does this article beyond noting that search engines will occasionally confuse your spreadsheet infrastructure with a soap opera platform.
The open question
The strategic risk for Univer is not technical. The team has already proven it can build a spreadsheet engine that people use; the architecture is coherent; the agent timing is good. The risk is that the agent-office category consolidates around something simpler. Most agent spreadsheet work today is bulk row processing and formula generation — tasks Autosheet handles inside the tools people already have, and Microsoft handles with Copilot inside the tool everyone has. The worktree-and-diff workflow is Univer’s most original contribution, but it presumes a world where agents produce office documents as first-class deliverables that humans review formally, rather than chat artifacts consumed on the fly.
If that world arrives — if “my agent opened a PR against the Q3 model” becomes a normal sentence — Univer has built for it earlier and more thoroughly than anyone else in open source. If it doesn’t, Univer remains a solid embeddable office SDK with a strong sheets engine, an honest license boundary, and a decade of hard-won spreadsheet scar tissue behind it. That is not a bad floor to fall back to.
Sources
- Univer: Office SDK & Office Harness for AI Agents
- Can the WPS Office SDK be used to embed document ...
- An open-source agent that edits your Word, PDF and Excel ...
- PDF to Office Conversion SDK
- AI Agent for Excel and Sheets | Autosheet (formerly GPT for ...
- Univer Video
- Web Office SDK Deployment: Hosted API to On-Premise
- 5 AI Agents That Replace Manual Spreadsheet Work
- Univer Video — Stream Movies & Live TV
- java - are there any ways to embed Open Office into my ...
- 8 Best AI Spreadsheet Automation Tools for Agents (2026)
- Univer Video: Filmes e Séries – Apps on ...