Univer is a spreadsheet and document SDK — the thing you embed when you want Excel-like editing inside your own product. It has 25,626 stars, an Apache-2.0 license, 121 releases since December 2023, and a README whose first line is “Build your own AI office with Univer Office SDK.”
That framing is doing a lot of work. An “office SDK” implies five products. Apache-2.0 implies you can ship all of it. Neither survives contact with the vendor’s own comparison page, and the part that does survive is better than the marketing: the open-source formula engine is genuinely strong, and the commercial boundary is not where the pricing page says it is.
I installed the packages, ran the headless engine in Node, called the commercial license module’s API directly, decompiled its watermark, and benchmarked both formula engines. Everything below is measured on the machine that wrote this article, not quoted from a landing page.
The measurements first
| Metric | Value |
|---|---|
| Stars / forks / watchers | 25,626 / 2,343 / 530 (2.07% of stars) |
| Open issues | 175 |
| Created | 2022-09-29 |
| Releases | 121, from 2023-12-20 to 2026-11-14 (v1.5.0) — 0.8/week |
| Release notes, median length | 5,847 characters (only 1 of 121 is under 50) |
| Contributors | 77 listed, 5,803 commits; top contributor 1,289 (22.2%); 60 of 77 have ≤8 commits |
| npm packages in one registry search | 91 @univerjs/* vs 153 @univerjs-pro/* |
@univerjs/engine-formula unpacked | 40.99 MB, 1,070 files |
@univerjs/engine-render unpacked | 28.8 MB, 491 files |
| Headless install footprint | 124 MB node_modules, 21 Univer packages |
| Features marked “Open source” on the vendor comparison page | 52 of 448 |
| Formula probes that compute headless | 36 of 40 |
The watcher ratio is the first thing worth noticing. 530 watchers on 25,626 stars is 2.07% — low in absolute terms, but far healthier than the “star as bookmark” pattern (0.28%–0.55%) I have measured on other recent audits. Weekly releases with 5,800-character release notes are the opposite of a code drop. This project is maintained.
That distinction — a paid boundary that moves rather than a product that never existed — is the same one I documented in the OpenMausBot audit, where the security sandbox switches on only in the hosted version. Univer’s version of it is subtler and, in my reading, more defensible: the open code is the engine, and the commercial layer is output polish.
What actually computes without a license
The README points you at a “Headless guide” for processing spreadsheets in Node without a browser. I followed it: @univerjs/core, engine-render, engine-formula, docs, sheets, sheets-formula, sheets-numfmt, no license key, no account, no network call.
Then I asked it 40 questions, one per cell, spaced six rows apart so dynamic-array spills could not collide:
| Function family | Probes | Result |
|---|---|---|
| Legacy aggregates and lookups | SUM, AVERAGE, VLOOKUP, INDEX-style | Pass |
| Modern lookups | XLOOKUP, XMATCH (incl. not-found default) | Pass |
| Dynamic arrays | FILTER, SORT, UNIQUE, SEQUENCE, VSTACK, HSTACK, TAKE, DROP, CHOOSECOLS, TOROW, TOCOL, WRAPROWS | Pass |
| Lambdas | LET, LAMBDA (=LAMBDA(x,x+1)(1) → 2) | Pass |
| Text | TEXTJOIN, TEXTBEFORE, TEXTSPLIT, CONCAT, TEXT with [$-409] locale format | Pass |
| Criteria | SUMIFS, COUNTIFS, IFS, SWITCH | Pass |
| Matrix | MINVERSE | Pass |
| Google-only | ARRAYFORMULA, GOOGLEFINANCE | #NAME? |
| Commercial-tier | SPARKLINE | #NAME? |
36 of 40. The four failures are three functions Univer never claimed (two are Google Sheets–only) and SPARKLINE, which the comparison page also does not mark as open source. I re-ran the identical probe against both the version the headless preset pins (1.0.4) and the current 1.5.0 line: 36 of 40 in both, same four failures.
That result matters because it is the honest case for Univer. If you need =XLOOKUP and =FILTER to recompute server-side on a snapshot, the Apache-2.0 packages do it. The comparison page’s own footnote concedes the limit is performance, not function: “The open-source edition includes only the frontend editors, not server components. Its formula engine has lower performance than the commercial edition.”
So I measured that too.
The performance claim, benchmarked
Two installs, same Node version, same generated workbook: 20,000 rows where column C is =An*2+Bn. One run with @univerjs/engine-formula, one with @univerjs-pro/engine-formula. Three runs each, median reported.
| Workload (20,000 formula cells) | Open source | Commercial engine | Difference |
|---|---|---|---|
| Flat two-operand formulas | 2,219 ms | 1,864 ms | ~16% faster |
Single dependency chain (=C(n-1)+1) | 1,795 ms | 1,366 ms | ~24% faster |
The commercial engine is faster. It is not a different class of engine: one commercial run on the flat workload took 2,722 ms, inside the open-source range. If the phrase “lower performance” made you expect an order of magnitude, it is 16–24% on synthetic 20,000-cell sheets.
The interesting measurement here is not the delta — it is that I could run the commercial engine at all, with no license, and get correct results.
The paid boundary is a watermark, not a gate
@univerjs-pro/license is the package that governs your commercial rights. It ships obfuscated: string-array indirection, _0x identifiers, the whole bundle (22 KB) run through a JavaScript obfuscator. Its type definitions, however, are plain, and its public README says this:
Front-end chart types, sparkline editing, pivot-table counts and dashboard creation are available with absent, restricted, valid or expired Licenses. License points no longer reduce these capabilities. Signed authorization still determines license watermarks… Local print and screenshot APIs also stay available regardless of License points or expiration; License affects their authorization watermarks rather than page counts or usage cooldowns.
The type definitions are blunter:
/** Front-end capabilities are open regardless of License points; watermark authorization is separate. */
export declare function isFeatureAvailableWithFreeTier(_accessor: IAccessor, _feature: IFeatureType): boolean;
/** Draw the mandatory license mark on independent output canvases, not free-tier permissions. */
export declare function renderLicenseWatermark(accessor, feature, canvas, logicalSize?): void;
Both parameters of isFeatureAvailableWithFreeTier are prefixed with _, which is how TypeScript marks an unused argument. I called it with a feature name that does not exist:
isFeatureAvailableWithFreeTier({}, 'sf') // true
isFeatureAvailableWithFreeTier({}, 'zzz-not-a-feature') // true
isFeatureAvailableWithFreeTier(null, undefined) // true
It returns true for everything. It is a constant with a suggestive name.
The watermark it draws is real, and I captured it by handing renderOutputWatermark a fake canvas whose 2D context records fillText calls:
const calls = [];
const ctx = new Proxy({}, { get: (t, p) => p === 'fillText'
? (s) => calls.push(String(s))
: p === 'measureText' ? (s) => ({ width: String(s).length * 7 })
: () => {} });
renderOutputWatermark(fakeCanvas);
// ["sales@univer.ai", "Contact Us: ", "Univer Office License Required"]
Three strings: a sales address, a contact prefix, and the phrase Univer Office License Required.
That phrase does not appear anywhere in the bundle as a string literal. It is assembled at runtime with charAt() calls over unrelated constants — LocaleType.EN_US, the default font stack, RENDER_CLASS_TYPE.BASE_OBJECT — so that grep -r "Univer Office License Required" on the package returns nothing. I reconstructed it by resolving the obfuscator’s string table and replaying the index pairs; the first character comes from "enUS".charAt(2).
That is a deliberate choice, and it tells you something the README does not: the vendor expects people to inspect the package, and prefers the free-tier watermark not be findable by grep.
What the license payload can still limit
The obfuscation hides the enforcement path, not the schema. lib/types/common/type.d.ts documents the signed payload your key decodes to:
| Field | Meaning |
|---|---|
dm | Domains allowed by the license |
rt | ReleaseType: DEV=1, NO_COMMERCIAL=2, COMMERCIAL=3 |
uf | Max units, max members, concurrent units, expiry |
sf | Pivot table count, max import size, max print number, reviews flag |
df | Doc import size, print count, reviews |
wsf | Max history number |
ReleaseType.NO_COMMERCIAL is the one to read twice. The license signature verifies (via @noble/ed25519 against a hardcoded public key, 69c4eb31…) whether your key permits commercial use at all — which is orthogonal to whether the code runs. The code always runs.
Three documents, three different answers about “freemium”
This is the part that will cost a team money if nobody reads it carefully, because all three statements are on the vendor’s own properties:
| Source | What it says freemium means |
|---|---|
| Repository README, “Scope and licensing” | “Freemium evaluation includes watermarks and feature limits.” |
| Pricing page | Freemium — “With watermarks and feature limits.” Links to “View limits”. |
@univerjs-pro/license README + util.d.ts | Feature points “no longer reduce these capabilities”; “Front-end capabilities are open regardless of License points.” |
Both marketing surfaces promise feature limits; the shipping code says features are open and only watermarks remain. I could not find a page that reconciles them, and I looked: the pricing page’s “View limits” link goes to documentation that does not enumerate a free-tier feature table.
The practical consequence is upside-down from what you would budget for. If you are building an internal tool, the open-source packages plus an unlicensed pro plugin will work — charts render, pivots compute, exports run — and your output will carry sales@univer.ai on it. The thing you are buying at $99/month is not capability. It is the absence of a watermark, plus server components.
This is the third documentation-versus-code gap I have hit this month in an “open core” project. The claude-mem audit found a 10,000-character ceiling that truncates without an error; Univer’s equivalent is a feature-limit table that exists in the type definitions, is called by nothing I could find, and is promised by two marketing pages. Both are the kind of defect that only shows up when you run the code instead of reading about it.
Measured from the pricing page: $99/month or from $1,000/year, covering 25+ editing connections (the page’s own conversion: ~1,000 MAU), 5,000+ exports/month, one domain and its subdomains, “No watermark · unlimited imports and printing.”
The comparison page disagrees with the README about Slides
The README says plainly: “current Slides and Boards capabilities are separately distributed commercial products.” The npm registry disagrees:
| Package | Scope | Latest | License field | Last publish |
|---|---|---|---|---|
@univerjs/slides | open | 1.3.1 | Apache-2.0 | 2026-10-30 |
@univerjs/slides-ui | open | 1.3.1 | Apache-2.0 | 2026-10-30 |
@univerjs-pro/slides | commercial | 1.5.0 | none | 2026-11-14 |
@univerjs/core | open | 1.5.0 | Apache-2.0 | 2026-11-14 |
Slides exists in both tracks. The Apache-2.0 one is real ("Core presentation model and services for Univer Slides", 178 KB unpacked, 26 files) and has been frozen at 1.3.1 while everything else moved to 1.5.0. The commercial one, first published 2026-06-04, is now the maintained line.
That is a legitimate open-core pattern — fork a capability, keep the old tree licensed — but it is not what the README says, and a team that read “Slides is commercial” and skipped it will never discover the Apache-2.0 1.3.1 packages sitting in the open scope.
Worse for planning: none of the commercial packages declare a license field at all. Not @univerjs-pro/license, not @univerjs-pro/sheets-chart, not @univerjs-pro/sheets-pivot. Neither does @univerjs/preset-sheets-advanced, the convenience preset that bundles them. Checked against the registry: license: undefined.
@univerjs/preset-sheets-advanced has 27 dependencies. All 27 are @univerjs-pro/*. So a package named in the open scope, published with no license metadata, resolves to exactly zero open-source code — while @univerjs/presets, one directory up, is Apache-2.0 with 9 dependencies and no pro packages at all.
Every pro package’s repository field points at https://github.com/dream-num/univer-pro, which returns 404. The bugs field points back at the public repo’s issue tracker. You can install the code; you cannot read it, file against it, or verify it — and npm audit-style provenance checks have no repo to compare against.
The version trap in the headless path
The Headless README says “Keep all @univerjs/* packages on the same version.” Its own package cannot:
// @univerjs/preset-sheets-node-core@1.0.4 dependencies
{ "@univerjs/core": "1.0.4", "@univerjs/sheets": "1.0.4", ... }
Exact pins, no caret. So following the headless guide installs the entire stack at 1.0.4 while the documentation, the comparison page, and every pro package are on 1.5.0 — published five minor versions later, 2026-10-10 vs 2026-11-14.
Then add the two pro packages a real integration needs (license and sheets-pivot) and npm cannot dedupe:
| Tree state | node_modules | Copies of @univerjs/core | Univer packages duplicated |
|---|---|---|---|
| Headless preset only | 124 MB | 1 | 0 |
+ @univerjs-pro/license + @univerjs-pro/sheets-pivot | 645 MB | 13 | 11 |
Thirteen copies of the core runtime — one at 1.0.4 from the preset, twelve nested at 1.5.0 under pro packages that require the newer version. Five copies of engine-formula. If you have ever debugged a plugin that “registered but did nothing,” duplicate injectors are the first thing to suspect, and this dependency graph manufactures them by default.
My recommendation if you go this route: skip the node-core preset, install the explicit package list from the Headless guide at a single version (1.5.0), and add overrides to force one copy. It is four lines and it removes an entire class of failure.
Where the open-source code ends, in the vendor’s own numbers
The comparison page counts features per product and tags some rows “Open source — Fully supported in the open-source edition.” I clicked through all seven tabs and counted the tagged rows:
| Product | Features listed | Rows tagged Open source | Share |
|---|---|---|---|
| Sheets | 83 | 36 | 43.4% |
| Docs | 73 | 16 | 21.9% |
| Slides | 66 | 0 | 0% |
| Canvas | 51 | 0 | 0% |
| Relational Tables | 62 | 0 | 0% |
| 70 | 0 | 0% | |
| Charts | 43 | 0 | 0% |
| Total | 448 | 52 | 11.6% |
Four of the five product surfaces in the “office SDK” have zero open-source features. Only two rows in the entire 448 are explicitly marked “Pro” (slicers and PivotTable Timeline, both v1.3.0+ · Pro) — everything else commercial is simply unlabelled, which is why the “Open source only” filter is the only reliable way to read that page.
The same page is candid about the shapes of what remains closed: no VBA execution, no XLSB, no ODS, no password-protected files, no Remove Duplicates, no Text to Columns, no Goal Seek, no chart sheets, no QUERY function, no iterative-convergence trace UI. XLSX import/export is “Supported” but untagged — consistent with the README, which puts Office file conversion in the Server SDK.
What to do with this
Univer is not a bait-and-switch. The Apache-2.0 core is real code that does real work, the release cadence is serious, and the vendor documents the commercial scope more honestly than most open-core projects. The problems are all recoverable, but only if you know about them before you build:
- Decide whether you’re shipping watermarks. Everything front-end works unlicensed. If your output goes to customers, budget $99/month from day one rather than discovering
sales@univer.aion a customer-facing PDF. - Pin one version, explicitly. Do not install
preset-sheets-node-coreand then add pro packages. Use the seven-package list from the Headless guide at1.5.0and add anoverridesblock. - Verify the license field before your legal review does. Commercial packages have no SPDX identifier in their published metadata. Get the terms in writing from sales; do not infer them from Apache-2.0 files in the public repo.
- Assume you cannot read the paid code.
univer-prois a 404 repo. If auditability is a requirement, treat every@univerjs-pro/*dependency as a binary-grade trust decision. - Test your own functions, not mine. My 40 probes are a floor.
=ARRAYFORMULAand=GOOGLEFINANCEfail by design; anything you actually depend on should be run headless against your real snapshot before you commit.
FAQ
Is Univer really open source?
The packages in dream-num/univer are Apache-2.0 and usable in commercial products. That covers content models, the command system, both engines, Facade APIs, the UI foundation, and Sheets and Docs editing. It does not cover charts, pivots, printing, collaboration, Office file conversion, or the server side.
Do I need a license key to use charts or pivot tables?
No. Those plugins import, register, and compute without a key — I verified the formula engine and the license module’s own isFeatureAvailableWithFreeTier returns true unconditionally. What the key removes is the output watermark.
What does the watermark look like?
Three canvas strings: sales@univer.ai, Contact Us: , and Univer Office License Required, drawn on output canvases (exports, screenshots) rather than the editor. It is applied by renderOutputWatermark, which is exported and callable directly.
Can I run its formula engine on a server with no browser?
Yes. @univerjs/core + engine-render + engine-formula + sheets + sheets-formula + sheets-numfmt compute correctly in Node with no canvas mounted. That install is 118–124 MB and gives you 36 of 40 functions from my probe set, including LET, LAMBDA and XLOOKUP.
How much does “office” cost after the free tier? $99/month or from $1,000/year for the Business plan (25+ editing connections, 5,000+ exports/month, no watermark, one domain plus subdomains). Enterprise is unlimited connections and exports at custom pricing.
Why 13 copies of @univerjs/core?
Because @univerjs/preset-sheets-node-core pins exact 1.0.4 dependencies while pro packages require 1.5.0. npm cannot dedupe two incompatible exact ranges, so it nests one copy per pro package branch. Installing at a single version removes it.
Bottom line
The gap between what Univer’s README promises and what the packages deliver is not the usual open-core gap. It is narrower in one direction — the free formula engine is better than “lower performance” implies, at 16–24% behind the paid one on 20,000 cells — and wider in another: 52 of 448 listed features carry the open-source tag, four of five product surfaces carry none, and the paid boundary is enforced by an obfuscated watermark whose text you cannot find with grep.
If your requirement is “recalculate and edit spreadsheets server-side, in my product, under a license I can read,” Univer’s Apache-2.0 core genuinely delivers that. If your requirement is “an office suite I can ship without paying,” read the pricing page’s smallest print twice: watermarks and feature limits is one claim, and the code implementing it says something else.