TL;DR

GenOffice is a free, Apache-2.0 licensed desktop office suite from Genspark (MainFunc.ai) that edits real Microsoft Office formats — .docx, .xlsx, .pptx — plus PDF and Markdown, with AI agents built into every editor rather than bolted on as a sidebar. It hit 3,566 GitHub stars and 597 forks in its first three weeks, and shipped v0.8.0 on August 23, 2026.

The single most important thing to know: most of the reviews you will find are already out of date. Coverage from the first week of August — including It’s FOSS and several aggregators — reported no Linux build, no Intel Mac support, and AI locked to Genspark’s credit system. All three of those limitations are gone as of v0.8.0.

Key facts:

  • 3,566 stars / 597 forks, repo created July 31, 2026 — roughly three weeks old
  • Apache-2.0, TypeScript, Electron, with a Rust sidecar for .xlsx I/O
  • Six apps, one engine layer: Docs, Sheets, Slides, PDF, Markdown, and a tabbed shell
  • Linux shipped: .deb, .rpm, and .AppImage (x86_64, glibc 2.34+), alongside signed macOS (Apple Silicon and Intel) and Windows installers
  • BYOK now standard — Claude, OpenAI, Gemini, DeepSeek, Kimi, GLM, Qwen, Doubao, MiniMax, Grok, Mistral, OpenRouter, or any OpenAI-compatible endpoint including local servers
  • Byte-preserving .docx round trip — only the paragraphs you touched are regenerated
  • Local PDF → Word/Excel/PowerPoint conversion, with system OCR for scanned pages on macOS and Windows
  • 40 open issues, alpha-grade in places — layout bugs and missing RTL support are real

Quick reference

Repogithub.com/genspark-ai/genoffice
Sitegenoffice.ai
LicenseApache-2.0
LanguageTypeScript (Electron) + Rust sidecar
Latestv0.8.0 — August 23, 2026
PlatformsmacOS 11+ (arm64 + x64), Windows 10+, Linux (glibc 2.34+)
InstallDownload installer from GitHub Releases
CostFree — no trial, no paid tier for the apps

What GenOffice actually is

Strip away the marketing and GenOffice is six Electron applications sharing a common set of TypeScript engine packages:

  • Docs — a .docx word processor built on Tiptap/ProseMirror
  • Sheets — an .xlsx spreadsheet built on the open-source Univer core, extended heavily in-house
  • Slides — a .pptx editor with an in-house parse/render engine
  • PDF — a viewer and true editor on pdf.js, pdf-lib, and PDFium wasm
  • Markdown — a block editor over plain .md files
  • Shell — the tabbed host with the home screen and auto-update

That structure matters more than it sounds. Most “AI office” products are a web app plus a chat panel that returns text you paste back into your document. GenOffice’s agents operate on the document model directly — block-level edits with version snapshots and diffs in Docs, and a tool-calling agent over workbook, slide, or PDF state in the others.

The project was, per It’s FOSS reporting, built largely by a single developer with about $10,000 of model tokens — which explains both the release cadence and the roughness in the corners.

Three things converged.

The format-fidelity claim. Open-source office suites have historically been judged on one axis: does the file still open correctly in Microsoft Office afterwards? LibreOffice’s answer is “usually, mostly, with caveats.” GenOffice makes a stronger and more specific promise, which we’ll dig into below.

The AI-office market is unusually polarized. Microsoft has pushed Copilot into Word, Excel, and PowerPoint whether users asked for it or not. Collabora and LibreOffice went the other way — AI as an optional add-on. GenOffice is the first credible entry that is simultaneously AI-first and fully open source, a combination that didn’t exist before August 2026.

The release velocity. Between August 16 and August 23 alone, the project shipped v0.7.204, v0.7.512, v0.7.686, v0.7.793, and v0.8.0 — five stable desktop builds in eight days.

The engineering that’s actually interesting: byte-preserving docx

Here is the part worth your attention if you’ve ever had an office suite mangle a document.

The standard approach to editing a .docx is: parse the whole file into an internal model, let the user edit, then serialize the entire model back out to OOXML. That serialization is lossy in practice — anything the parser didn’t fully understand (an obscure style, a content control, a custom XML part, a field code) gets dropped or rewritten. Open the result in Word and the layout has quietly shifted.

GenOffice inverts this. From the README’s architecture notes:

open docx ─► archive original by hash (never touched)
          ─► docx-engine parses word/document.xml top-level elements (w:p / w:tbl / …)
          ─► Block tree, each block anchored by docxIndex + original XML slice
          ─► Tiptap streaming editor (manual + AI editing, dirty tracking)
save      ─► dirty blocks → OOXML fragments (referencing existing styles only)
          ─► splice into original document.xml (untouched blocks keep original bytes)
          ─► repack zip; all other entries copied byte-for-byte

The original file is the source of truth. Each block keeps a pointer (docxIndex) back to its original XML slice. On save, only dirty blocks are regenerated into OOXML fragments and spliced back in — every untouched paragraph, and every other entry in the zip archive, is copied byte-for-byte.

The practical consequence: if you open a 40-page contract and change one sentence, the other 39 pages are bit-identical to what your colleague sent you. Word cannot detect that a different program touched them, because in a meaningful sense nothing did. The same philosophy is applied in Sheets and Slides — narrow patches, original file preserved.

This is a genuinely better architecture than the parse-everything-reserialize approach, and it’s the strongest technical argument for the project.

PDF editing that isn’t a cover-up annotation

The second engineering claim worth calling out is PDF text editing. Most PDF “editors” fake it: they draw a white rectangle over the old text and place a new text box on top. Print it, search it, or open it in another reader and the deception shows.

GenOffice rewrites the actual page content stream through PDFium wasm, with subset-embedded fonts and original-font preservation, doing paragraph selection with in-block reflow and alignment restoration. That’s the real thing.

It also does local PDF → Office conversion via packages/pdf2docx: PDFium character-level extraction, pure-geometry layout analysis, then rebuild through the docx engine. No cloud, no upload. Scanned pages are handled too — v0.8.0 added OCR through the operating system’s own engine on macOS and Windows, so image-only PDFs convert to editable text with nothing extra to install.

If you have ever paid a subscription to a web service purely to turn a PDF into a Word file, this feature alone is the reason to look at GenOffice.

Getting started

Download the installer for your platform from GitHub Releases. The macOS and Windows installers are signed with the Genspark company certificate.

On Debian or Ubuntu:

sudo apt install ./genoffice_0.8.0_amd64.deb

On Fedora, RHEL-family, or openSUSE:

sudo dnf install ./genoffice-0.8.0.x86_64.rpm     # Fedora / RHEL
sudo zypper install ./genoffice-0.8.0.x86_64.rpm  # openSUSE

The AppImage runs in place, but needs the FUSE 2 runtime:

sudo apt install libfuse2        # libfuse2t64 on Ubuntu 24.04
chmod +x GenOffice-0.8.0.AppImage
./GenOffice-0.8.0.AppImage

Building from source is a standard npm workspace setup:

npm install
npm run fixtures     # generate test .docx fixtures
npm test             # engine + app unit tests
npm run dev          # all five editors + shell via Vite

Sheets needs a Rust toolchain (cargo on PATH) for its xlsx sidecar; npm run build -w @genoffice/sheets compiles it automatically.

Configure BYOK first. By default the apps sign in to a Genspark account via a device-code flow and route model calls through Genspark’s proxy. If you’d rather not do that — and if you handle anything confidential, you shouldn’t — go into AI settings and drop in your own key, or point the custom provider slot at a local OpenAI-compatible server.

What the community is saying

Sentiment splits cleanly along two lines.

The architecture gets respect. The byte-preserving docx patching and real PDF content-stream editing are consistently singled out as legitimately good engineering, not marketing. Open Source For You framed it as embedding AI directly into native files “to eliminate sidebars and layout corruption.”

The AI plumbing got the criticism — and drove the actual roadmap. The most-discussed open issue on the repo is bluntly titled “Other AI integration?” (9 comments), followed by a detailed request for “User-owned AI configuration: custom model endpoint, Tavily search, proxy, agent rules/skills.” The early complaint was exactly what It’s FOSS flagged: AI features drew on Genspark’s credit system, tied to an account rather than a local model, with text passed to third-party providers under Genspark’s privacy policy.

That pressure worked. BYOK across twelve providers plus arbitrary OpenAI-compatible endpoints is now a headline feature. This is a maintainer who ships in response to issues, fast — which is the single best predictor of whether a three-week-old project is worth betting on.

The Linux question followed the same arc. It’s FOSS found the Linux build issue open with no maintainer reply and concluded “I wouldn’t hold my breath.” Three weeks later there are .deb, .rpm, and .AppImage artifacts in every release.

Honest limitations

This is a three-week-old alpha. Treat it accordingly.

  • Layout bugs are real. There is an open issue titled “[Docs] Layout breaks when opening .docx”. The byte-preserving save protects untouched content, but the rendering of complex documents is still maturing.
  • No RTL support. A comprehensive right-to-left support request for Arabic across the suite is open. If you work in Arabic, Hebrew, Persian, or Urdu, this is a hard blocker today.
  • No web version. Asked and open, no commitment.
  • Analytics on by default. Official packaged builds send limited usage analytics unless you disable it under Settings → General. Genspark states analytics never includes document content, file names, paths, account identity, or email — but you should know it’s on and turn it off if you care.
  • AI features need network. Document editing is fully local. Agents, search, and image tools are not — unless you point BYOK at a local model server, which the custom provider slot does allow.
  • Adoption is early. v0.8.0 has roughly 558 downloads across all six artifacts. Windows leads at 347. You will be an early user finding bugs.
  • Linux is x86_64 only. No ARM Linux build, and glibc 2.34+ rules out older LTS releases like Ubuntu 20.04.

GenOffice vs the alternatives

GenOfficeLibreOfficeOnlyOfficeMicrosoft 365
LicenseApache-2.0MPL-2.0AGPL-3.0 / proprietaryProprietary
CostFreeFreeFree tier + paidSubscription
OOXML fidelityByte-preserving patchGood, reserializesVery goodNative
AI modelAI-first, BYOKOptional add-onOptional add-onCopilot, bundled
Bring your own keyYes, 12+ providersN/ALimitedNo
True PDF text editYes (PDFium)Draw-basedLimitedLimited
Local PDF → OfficeYes, on-deviceNoServer-sideCloud
MaturityAlpha (3 weeks)25+ years10+ yearsDecades
Linuxdeb/rpm/AppImageExcellentExcellentWeb only

Choose LibreOffice if you need a mature, stable, deeply localized suite and don’t care about AI. It remains the correct default for most people.

Choose OnlyOffice if OOXML compatibility is your top priority and you want a polished, collaborative, proven product.

Choose GenOffice if you specifically want AI editing inside real Office files, on your own API key, on your own machine, with source you can read — and you can tolerate alpha rough edges. It is also the strongest free option today for local PDF-to-Office conversion.

Who should use this

Good fit: technical users who want an auditable AI office suite; anyone doing regular PDF-to-Word conversion on-device; people who want AI document editing without a Microsoft subscription or a Genspark account; contributors looking for a fast-moving TypeScript project with a responsive maintainer.

Bad fit: anyone whose work is mission-critical and deadline-bound today; RTL language users; teams needing real-time collaborative editing; organizations with formal software-approval processes. For those, wait for 1.0.

FAQ

Is GenOffice really free? Yes. Apache-2.0, no trial, no paid tier for the apps themselves. The only cost is model inference — either through a Genspark account or your own API key under BYOK.

Does GenOffice work on Linux? Yes, as of the v0.7.x/v0.8.0 releases. There are .deb, .rpm, and .AppImage builds for x86_64 with glibc 2.34+ (Ubuntu 22.04+, Fedora 35+, RHEL 9+, openSUSE Leap 15.6+). Articles published in early August 2026 saying otherwise are out of date.

Can I use my own API key instead of a Genspark account? Yes. BYOK supports Claude, OpenAI, Gemini, DeepSeek, Kimi, GLM, Qwen, Doubao, MiniMax, Grok, Mistral, and OpenRouter out of the box, plus a custom provider slot for any OpenAI-compatible endpoint — local model servers included.

Will GenOffice break my Word documents? Saving is byte-preserving: only paragraphs you actually edited are regenerated, everything else in the zip is copied byte-for-byte. That’s a stronger guarantee than most open-source suites offer. Rendering of complex documents is still buggy in places, so verify important files before sending them on.

Does GenOffice work offline? Document editing — opening, editing, saving .docx, .xlsx, .pptx, PDF, Markdown — is fully local and works offline. Local PDF-to-Office conversion and OCR are also on-device. AI agents, web search, and image generation need network access, unless you configure BYOK against a local model server.

Can it convert scanned PDFs to editable Word files? Yes, since v0.8.0. Scanned pages are read using the operating system’s own OCR engine on macOS and Windows — nothing uploaded, nothing extra to install. Linux OCR is not covered by that system-OCR path.

Does it collect my data? Packaged builds send limited usage analytics by default, disableable under Settings → General. Genspark states analytics excludes document content, file names, paths, account identity, and email. If you use the default Genspark sign-in rather than BYOK, AI prompt text is processed by third-party model providers under Genspark’s privacy policy — so use BYOK for confidential work.

The bottom line

GenOffice is three weeks old and it shows. But the two things nobody else does well — byte-preserving OOXML editing and genuine on-device PDF-to-Office conversion with OCR — are the hard parts, and they appear to be solved. The easy parts, like polishing layout rendering and adding RTL, are what’s still missing.

The project also has a demonstrated pattern: users filed issues about vendor lock-in and missing Linux support, and within three weeks both shipped. That responsiveness is worth more than the current feature list.

Watch this one. Don’t run your quarterly board deck through it yet.

Sources

Repo state and release data verified against the GitHub API on August 24, 2026.