Ask an engineer to let Agents Edit Cells in a spreadsheet and the reflexive shopping list starts with an AI SDK: wire up a model, bolt on tool orchestration, and only then touch the workbook. The dream-num/univer repository argues the opposite. See dream-num/univer; the plain address is https://github.com/dream-num/univer . This open-source Office SDK wraps workbooks, worksheets, and ranges into a set of Facade APIs, so an application or an Agent can write values, watch formulas recalculate, and persist the result as a JSON snapshot, all Without an AI SDK. The model and the tool orchestration stay in your own application. The Univer Office SDK simply puts Office capabilities inside your product.
As of November 15, 2026, reading the GitHub API cumulative totals, the repository carries a stargazers_count of 25192, a forks_count of 2307, TypeScript as its language, dev as its default branch, and a created_at of 2022-09-29. Its status is public and its license.spdx_id is Apache-2.0. Two readings deserve emphasis. The star figure is a cumulative total read from the stargazers_count field, not a single-day increment. The pushed_at field reads 2026-11-15, which is the moment these fields were fetched, not a release date, so it should not be paraphrased as a launch on that day. Every number here comes straight from the repository page and the GitHub API. This article checks only what the README, the LICENSE, and the GitHub API can confirm, and it adds no pricing, no benchmarks, and no competitor comparisons on the project's behalf, nor does it treat media retellings as a primary source.
Start with what this is. Univer is an Office SDK for building Office applications inside your own product, not a finished office suite an end user opens to write a document. The English README says it provides Office editing, calculation, collaboration, and file processing for people and existing agents, using your data and permissions. The repository description is blunter: Build your own AI office with Univer Office SDK, bringing editing, calculation, collaboration, and file processing into your product for people and AI agents on an extensible open-source core. The product you are building is your own office, and Univer supplies the embeddable parts.
Facade APIs Turn Low-Level Engines Into Application Objects
Facade APIs are the front door. The Chinese README calls them the application-facing interface to workbooks, worksheets, and ranges, wrapping low-level capabilities into objects such as FWorkbook, FWorksheet, and FRange. The English line is Facade APIs are the application-facing interface to workbooks, worksheets, and ranges. Why bother with the layer at all? Content models, editing, rendering, and formula calculation are awkward to expose directly to business code: they are fine-grained, tightly coupled, and a change in one corner ripples through the rest. Facade APIs reshape them into objects an application understands. An FWorkbook is one workbook, an FWorksheet is one sheet inside it, an FRange is a span of cells, getRange fetches that span, setValues writes a block of values into it, and createWorkbook rebuilds a workbook from a saved snapshot. Business code does not need to know how rendering refreshes or how the formula engine schedules its work; it only faces these handful of objects.
Underneath sits the foundation. The README states Web SDK is the foundation: content models, editing, rendering, formula calculation, and Facade APIs (application-facing APIs). Presets compose plugins and register their Facade APIs. That detail matters more than it first appears. An Agent does not take a side door or a special AI channel; it shares the very same Facade APIs as the editor and your application. The section heading Connect applications and agents through Facade APIs and events names exactly this arrangement, where applications and agents attach to one workbook through Facade APIs and events.
What the open-source core covers is spelled out in the README capability list, and these are self-described features rather than third-party measurements. For Docs it lists Rich-text editing, paragraph formatting, lists, hyperlinks, comments, and drawing integration. For Sheets, under Editing and layout, it lists Cell styles, merged cells, in-cell rich text, freeze panes, floating images, and undo/redo. Cell styles, merged cells, in-cell rich text, frozen panes, floating images, and undo and redo are capabilities aimed at real editing sessions rather than a demo-only kernel. The docs landing page notes that LLMs and Agents call Facade APIs directly for automated document reading, writing, and analysis.
Agents Edit Cells Without an AI SDK
Here the contrast in the title lands. Letting Agents edit cells does not require buying an extra AI SDK. The README example is deliberately small: sheet.getRange('A1:D4').setValues([...]) writes values into a block, and sheet.getRange('B2').setValue(4) changes one cell to 4, with the comment // An application control or agent can make the same edit. Whether the caller is an application control or an agent, it is the same edit, with no second interface and no dedicated AI pathway. Set out on its own, the example looks like this:
sheet.getRange('A1:D4').setValues([...])
// An application control or agent can make the same edit.
sheet.getRange('B2').setValue(4)The official text then adds the key premise: The editor and your application operate on the same workbook; formula changes are recalculated automatically. Once an Agent writes B2, the formulas depending on it do not need a manual refresh, because the calculation engine follows on its own. The person at the screen, the logic in the background, and the running Agent all see one consistent result.
So where does the intelligence come from? The README draws that line clearly: Agents can use the same open-source APIs; your application supplies the model and tool orchestration. Agents reuse the same open-source APIs, while the model and the orchestration live in your application. The SDK ships no built-in model and bundles no Agent orchestration framework. It guarantees only that a human, application logic, or an Agent all land on the same Facade APIs. The next sentence states the thesis of this article: The quotation example demonstrates editing, calculation, and saving without an AI SDK dependency. The quotation example walks through editing, calculation, and saving Without an AI SDK. That is the particular move of the Univer Office SDK for Agents: it leaves the AI on your application side and keeps the Office inside the library, so neither side has to package the other.
workbook.save() Is a JSON Snapshot, Not an XLSX File
The second common misreading concerns saving. workbook.save() does not return an XLSX file; it returns a JSON snapshot. The English README is precise: workbook.save() returns a JSON snapshot containing the workbook and its registered plugin data. Your application owns persistence; this example uses browser storage. The snapshot carries the workbook itself plus the data of its registered plugins, and where it goes, how it is stored, and how long it lives are decisions for your application. The line localStorage.setItem('quotation', JSON.stringify(workbook.save())) is only a browser-storage demonstration, not a prescribed persistence layer. Restoring reverses the step, as the paired example shows:
localStorage.setItem('quotation', JSON.stringify(workbook.save()))
if (saved) { univerAPI.createWorkbook(JSON.parse(saved)) }Here createWorkbook rebuilds the workbook from the JSON snapshot. Store and restore close the loop inside your own application, and the SDK does not pick your storage backend for you.
Keep that path separate from XLSX import and export. The README says This workflow uses only open-source packages. Snapshot persistence is separate from XLSX import and export, which uses the separately distributed file-conversion capabilities. Snapshot persistence runs on open-source packages, while XLSX import and export rely on separately distributed file-conversion capabilities; they are not the same thing. Treating workbook.save() as an Excel export, or assuming the open-source core ships complete XLSX reading and writing, flattens the licensing boundary. What you get is a JSON snapshot of workbook state plus createWorkbook to rebuild it, while file-format conversion sits on a different distribution line.
Nor is the snapshot limited to the browser. The official Node.js documentation says In Node.js, you can create or load workbooks, read and write cells and formatting, change rows and columns, calculate formulas, and save Univer snapshots. Creating or loading workbooks, reading and writing cells and formatting, changing rows and columns, calculating formulas, and saving Univer snapshots all work server-side. For a backend that batches spreadsheet work and lets the application supply its own model and orchestration, this Node.js path is the same Facade APIs logic in a different runtime.
Where Apache-2.0 Stops: The Open-Source Boundary
The licensing boundary deserves a word-for-word reading. The English README states: This repository provides the Web SDK's core architecture and engine implementations under Apache-2.0: content models and snapshot formats, plugin and dependency management, the command system, formula and rendering engines, Facade APIs, UI foundation, and Sheets and Docs editing capabilities. The Chinese README matches it: the repository provides the Web SDK core architecture and engine implementations under Apache-2.0, while commercial Web SDK plugins, Server SDK, and commercial AI SDK extensions are distributed and licensed separately. The scope is explicit. Apache-2.0 covers the Web SDK core architecture and engine implementations, from content models and snapshot formats through plugin and dependency management, the command system, formula and rendering engines, Facade APIs, the UI foundation, and Sheets and Docs editing capabilities. It is not a blank promise that everything in the repository is free to take.
The second half completes the picture: Commercial Web SDK plugins, Server SDK, and commercial AI SDK extensions are distributed and licensed separately. These commercial plugins, the Server SDK, and the commercial AI SDK extensions are distributed and licensed on their own line. The Chinese license section adds Copyright (c) 2021-present DreamNum Co., Ltd., notes that the open-source packages in this repository use Apache-2.0 while commercial SDK modules are distributed and licensed separately, and mentions that Freemium evaluation includes watermarks and feature limits. The reading closest to the text is that Apache-2.0 grants the Web SDK core architecture and engine implementations, while the commercial modules sit outside that grant. Mistaking the localStorage demo for a persistence layer, treating the JSON snapshot as an XLSX file, or counting commercial plugins inside Apache-2.0 all come from erasing that boundary.
Stack the four layers and the position of this Univer Office SDK is clear. Behind 25192 cumulative stars is not another chat-flavored office toy but a set of parts that push editing, calculation, collaboration, and file processing into your own product. Facade APIs put FWorkbook, FWorksheet, and FRange in front of the application; getRange and setValues let Agents edit cells as directly as changing a variable; workbook.save() hands off state as a JSON snapshot and createWorkbook rebuilds it; and the Apache-2.0 boundary stops at the Web SDK core and engines, with commercial Web SDK plugins, Server SDK, and commercial AI SDK extensions on a separate line. None of that needs an AI SDK. This article is about the SDK project itself, a different angle from the workflow pieces ai-spreadsheet-sop and ai-agent-office-automation-sop on doing spreadsheet work with AI, and from the runtime comparison in agent-runtime-sdk-comparison-review.
FAQ
Q1: Must an Agent have a separate AI SDK before it can edit cells? A1: No. The README says Agents can use the same open-source APIs, so an Agent reuses the very same Facade APIs as the editor and your application. getRange and setValues change cells directly, the model and tool orchestration come from your application, and the quotation example runs Without an AI SDK.
Q2: Does workbook.save() export an XLSX file? A2: No. workbook.save() returns a JSON snapshot containing the workbook and its registered plugin data, persistence belongs to your application, and the localStorage call is only a demo. XLSX import and export rely on separately distributed file-conversion capabilities, separate from snapshot persistence, while createWorkbook rebuilds the workbook from the snapshot.
Q3: Is every feature in the repository covered by Apache-2.0? A3: No. Apache-2.0 covers the Web SDK core architecture and engine implementations. Commercial Web SDK plugins, Server SDK, and commercial AI SDK extensions are distributed and licensed separately, and the Chinese license section notes that Freemium evaluation includes watermarks and feature limits.
Q4: Do formulas recalculate after an Agent changes a cell? A4: Yes. The editor and your application operate on the same workbook, and formula changes are recalculated automatically, so a changed value needs no manual refresh because the calculation engine follows.
Q5: Is 25192 a single-day gain in stars? A5: No. As of November 15, 2026 it is the cumulative stargazers_count total read from the GitHub API, not a single-day increment; the pushed_at value of 2026-11-15 is the moment the fields were fetched, not a release date.
Related reading
- ai-spreadsheet-sop: a workflow for handling spreadsheets with AI.
- ai-agent-office-automation-sop: steps for Agent office automation.
- agent-runtime-sdk-comparison-review: a comparison of runtime SDKs.
- invoice-ocr-to-excel-sop: a workflow from invoice recognition to Excel.
- officecli-resource: Office tooling resources for the command line.