GenOffice: An Open Source AI Office Suite

Genspark open sourced GenOffice on 3 August 2026, a desktop office suite with AI editing built into every document type. The code is on GitHub under Apache 2.0. It ships four apps, Docs, Sheets, Slides and PDF, reads and writes .docx, .xlsx, .pptx, .pdf and Markdown, and runs as a native application on macOS, Windows and, since a follow-up release, Linux, rather than in a browser tab.

The detail that got the most attention is the build story. Genspark says the GenOffice alpha was built by one engineer, in one week, for about $10,000 in model tokens. For readers of this blog that is the actual news. The office suite is the artefact; the economics are the story.

What is in the release

From the project's GitHub repository, the suite is:

• GenOffice Docs, a .docx word processor

• GenOffice Sheets, a .xlsx spreadsheet

• GenOffice Slides, a .pptx presentation editor

• GenOffice PDF, a viewer and editor

• GenOffice shell, a launcher that hosts the four in tabs

The README describes byte-preserving round trips, meaning parts of a file you did not edit come back unchanged. Anyone who has watched a converter mangle a colleague's tracked changes will understand why that is called out. The editing surface is built on Univer, an existing Apache-2.0 spreadsheet and document core, with Genspark's own layer on top. The suite launched at version 0.4.110 labelled alpha; the project has since moved past 0.5.0 across platforms, and its README no longer uses the word alpha, though that is a description to verify yourself rather than take as a maturity guarantee.

The AI layer is not a chat sidebar bolted on. In Docs it does block-level editing with version snapshots and diffs, so you can see what the model changed and revert a specific block. In Sheets, Slides and PDF it runs as a tool-calling agent over the document state.

The catch worth reading carefully

Apache 2.0 covers the code. It does not cover the intelligence.

The apps sign in to a Genspark account and route model calls through Genspark's service. No model API key is stored locally, which sounds like a convenience and is also a dependency: the editing works offline, the AI does not, and the AI is metered by a company that can change its terms. There is also a reserved ee/ directory in the repository for enterprise modules under separate licensing, which is a normal open-core arrangement and worth knowing about before you plan around it.

Update since publication: Genspark added bring-your-own-key (BYOK) support in the weeks after launch. The README now lists Claude, OpenAI, Gemini, DeepSeek, Kimi, GLM, Qwen, Doubao, MiniMax, Grok, Mistral, OpenRouter, or any OpenAI-compatible endpoint as usable in place of Genspark sign-in, with the Genspark account as the default rather than the only path. That does not remove the dependency described above for anyone who accepts the default, but it does mean the single-vendor lock-in is now something you can configure your way out of, not something built into the license.

So "free and open source AI office" is accurate about the office and, since the BYOK update, closer to accurate about the AI too, for anyone willing to configure it. If you leave the default Genspark sign-in in place and Genspark changes pricing or shuts down, you keep a capable offline editor and lose the feature the product is named for. Switch to your own API key first and that risk mostly goes away, which is a materially different, and better, risk profile than the suite shipped with at launch.

Why the build story matters more than the suite

One engineer, one week, $10,000 of tokens, and the result is a four-application desktop suite with file-format fidelity. Take the claim at face value for a moment. It does not mean office suites are now trivial. It means the cost floor for producing a credible, specific piece of software has moved somewhere most planning assumptions have not caught up with.

Two things made it possible, and neither is magic. The team built on Univer rather than writing a spreadsheet engine, so the hardest part was already solved and permissively licensed. And the work is exactly the sort of well-specified, heavily-precedented software that current models handle best: known file formats, known UI patterns, an existing core to integrate against.

That is the useful lesson. The projects where AI-assisted development collapses timelines are the ones with a clear specification and existing components to lean on. The projects where it does not are covered in the limitations of AI app builders. If you are weighing a build of your own, how to build an app with AI is the practical starting point, and what an AI coding agent actually does explains the tooling that makes a week-long build plausible.

One more thing to settle before you copy the model: a codebase written mostly by a model raises real questions about who holds the copyright, which we cover in who owns AI generated code. If you are weighing whether to release your own model-built project this way, should you open source your AI side project is worth reading first.

FAQ

Is GenOffice really free?

The application is, under Apache 2.0. The AI features work by default through a Genspark account and Genspark's service, but the project added BYOK (bring your own key) support after launch, so you can point the AI panel at your own Claude, OpenAI, Gemini or other API key instead.

Can I use GenOffice without an internet connection?

The editors are native desktop applications and work offline. The AI panel calls out to Genspark's service, so it does not.

Does GenOffice work with Microsoft Office files?

Yes. It reads and writes .docx, .xlsx and .pptx, and the project claims byte-preserving round trips for parts of a document you have not edited.

Can I plug in my own model?

Yes, as of a post-launch update. GenOffice added BYOK (bring your own key) support for Claude, OpenAI, Gemini, DeepSeek, Kimi, GLM, Qwen, Doubao, MiniMax, Grok, Mistral, OpenRouter, or any OpenAI-compatible endpoint, configurable in settings as an alternative to Genspark sign-in. It was not there at launch, so check your installed version if this matters to you.

Is it ready for real work?

It launched describing itself as alpha at version 0.4.110. The project has since moved past version 0.5.0 across macOS, Windows and Linux, and current README copy no longer uses the word alpha. Treat maturity claims as something to verify against your own files rather than take on faith, and still evaluate it on copies before trusting it with documents your business depends on.

For a look at another lightweight way AI can extend everyday software, see can AI build a Chrome extension which covers what AI handles well and what still needs a human.