An Office Suite Built in a Week for $10K. The AI Phones Home.

TL;DR

Genspark open-sourced GenOffice, a five-app native office suite covering .docx, .xlsx, .pptx and PDF, under Apache License 2.0. The repo went public on July 31 and the first tagged build, v0.4.110, landed August 2. The company says one engineer built the alpha in a week on $10,000 of tokens. The file engines are the interesting part, and the licensing is the part you should read before you get excited: AI features sign in to a Genspark account and proxy through Genspark's servers, and the single non-Apache directory in the tree is reserved for private deployment.

What actually shipped

Five Electron apps over one shared TypeScript engine layer: Docs, Sheets, Slides, PDF, and a shell that tabs the other four and handles auto-update. Signed installers exist for macOS on Apple Silicon and Windows x64. It is TypeScript throughout, with one Rust sidecar.

The dependency choices are unusually honest for a launch like this. Sheets is built on the open-source Univer core with a large in-house extension layer, and its xlsx import and export run through a Rust sidecar wrapping calamine and IronCalc. PDF is pdf.js plus pdf-lib. Docs runs a streaming TipTap editor. Slides is the one genuinely in-house engine, with HarfBuzz metrics for text shaping.

Genspark calls this the world's first full-featured open-source AI office. LibreOffice might have notes on the "first open-source office suite" framing. The load-bearing adjective is "AI-native," and on that claim the architecture backs them up.

The docx round trip is the real engineering

Most editors that touch a .docx parse it into their own model and regenerate the entire file on save. That is why a document round-tripped through a third-party editor comes back with the margins subtly wrong, the numbering restarted, and a style you never asked for.

GenOffice archives the original file by hash and never touches it. It parses the top-level elements of word/document.xml into a block tree, where every block keeps an index anchor and a slice of the original XML. On save, only dirty blocks are regenerated into OOXML fragments, those fragments are spliced back into the original XML, and every other zip entry is copied byte for byte.

Think of it as repairing a book by replacing the one page you rewrote, rather than retyping the whole manuscript and hoping the pagination survives. Sheets and Slides follow the same rule: the original file is the source of truth and edits are narrow patches.

This matters more than it sounds when an agent is doing the editing. An LLM rewriting your quarterly report is a much smaller risk if the blast radius of a bad edit is three paragraphs of XML instead of the entire document.

The part where it phones home

Here is the sentence in the README that decides whether this is useful to you: the apps sign in to a Genspark account and route model calls through the Genspark service side, with no model API key stored locally. The provider types file goes further and describes Genspark login as "the sole auth source for AI features."

Read the code and it is more nuanced. providers.ts ships six providers, including direct Anthropic, Gemini, DeepSeek and OpenAI entries with real API key placeholders, plus an OpenAI-compatible "custom" provider with a base URL field. The security policy confirms requests are proxied through the signed-in account by default while user-supplied keys stay in the OS settings store.

The plumbing for bring-your-own-model is there. The settings page to reach it is not. Within a day of launch someone filed issue #7 asking whether Gemini, Grok or a local LLM could stand in for Genspark, and a commenter noted you would have to edit the provider file by hand to do it today.

That last detail is worth staring at. The Genspark proxy list includes Claude Opus 4.7 and 4.8, GPT-5.2 and Gemini 3.1 Pro preview, with Claude Opus 4.7 as the default. The direct OpenAI entry you would use with your own key tops out at GPT-4.1 and GPT-4o. Bringing your own key is technically supported and quietly a downgrade.

The empty directory that explains the business model

The repo is Apache 2.0 with exactly one carve-out: the ee/ directory, covered by a separate GenOffice Enterprise License held by Mainfunc, Inc. Production use of anything in there requires an enterprise agreement.

The enterprise directory currently contains two files: a license, and a README explaining the license. That README is unusually candid about what is coming, naming "private deployment and offline license verification" as the reserved modules.

So the feature that would let you run this thing without calling Genspark's servers is, by design, the paid one. That is a defensible open-core split and the maintainers are being upfront about the boundary, which is more than most. But if you came here because "open-source AI office suite" sounded like something you could point at a local model on your own hardware, the roadmap says otherwise.

Also: no Linux

Installers cover macOS and Windows only. In issue #4 a maintainer confirmed there is no technical blocker, just product prioritization, and that an AppImage plus a Linux build of the xlsx sidecar is planned once the feature set settles. CI already runs on ubuntu-latest, so Linux can build GenOffice. It just cannot download it.

The security posture is better than you would expect

For a one-week alpha, the hardening is real. Every window runs with contextIsolation on, nodeIntegration off and sandbox on. IPC payloads are schema-checked in the main process, with zod end to end in Sheets. External link opening goes through a single gate with a protocol allowlist that rejects file:, javascript: and custom schemes.

The most interesting bit is how Slides handles AI-generated layout changes. The model emits what looks like a small JavaScript subset, but it is parsed with Acorn and executed by a constrained AST interpreter, never handed to eval, Function, a VM context or a worker. There are no ambient globals, no module loader, no network, no timers, and there are statement and call-depth limits. Every edit primitive validates its arguments and writes into an op buffer applied through the same command pipeline as manual edits.

Caveats before you install it

• It is alpha, and the bug reports match. Within hours of launch, users filed issues reporting that Save As on macOS mutates the original file before creating the copy, and that Save As on an xlsx can fail with a sheet not found in workbook.xml.
• Traction is early: 827 stars and 111 forks at the time of writing, one tagged release, and a handful of open issues.
• AI features require a Genspark account. There is no offline or fully local mode today.
• Forks must rebrand. The Apache 2.0 grant does not cover the GenOffice or Genspark trademarks.

Key Takeaways

• Genspark released GenOffice under Apache 2.0 on July 31, with v0.4.110 tagged August 2: five Electron apps covering docx, xlsx, pptx and PDF on macOS and Windows.
• The byte-preserving round trip is the standout engineering. Only edited blocks are regenerated and spliced back into the original file, so agent edits cannot silently reflow a document.
• AI calls default to a Genspark login and Genspark's proxy. Direct provider and OpenAI-compatible custom entries exist in the code, but there is no settings UI to reach them yet.
• The proxy is wired to Claude Opus 4.7 and 4.8, GPT-5.2 and Gemini 3.1 Pro preview. The direct OpenAI list stops at GPT-4.1 and GPT-4o, so bring-your-own-key is currently a step down.
• The one non-Apache directory, ee/, is empty except for a license and a README reserving private deployment and offline license verification as enterprise features.
• Genspark's claim that one engineer built the alpha in a week on $10,000 of tokens is the company's own, and it is the number everyone will repeat. The dependency choices and the security doc suggest more prior art and more care than that framing implies.

Sources: genspark-ai/genoffice on GitHub, GenOffice v0.4.110 release, packages/ai-provider/src/providers.ts, GenOffice SECURITY.md, ee/ enterprise license and README, issue #7: other AI integration, issue #4: Linux support, Genspark's launch announcement