Somebody submitted Orkas through our tool form last week. Most of what comes through that form is off-topic and gets a polite no, but this one is squarely the kind of thing we cover, so I spent a morning verifying what it actually is. Disclosure up front: it arrived as a free submission, nobody paid anything, and Orkas has no idea I'm writing this.
Orkas is an open-source desktop app that runs a team of AI agents on your own machine. A Commander model takes your goal and farms the work out to specialist agents: DeepResearcher for sourced research, ContentWriter, PptMaker for actual editable .pptx files, OfficeWorker for Word and Excel, VideoStudio, a UI designer, an SEO auditor. Nine ship installed, and there's a marketplace that claims thirty.
The pitch is the same one every agent product makes in 2026. What made me look closer is the stuff that's checkable.
The repo is real and actively worked on. github.com/Orkas-AI/Orkas, MIT license, 1,796 stars when I checked on September 7. Their own site says "1.6k+", which means their marketing understates their numbers. That's rare enough that I noticed. Eight releases between July 6 and August 29, the latest one improving Commander reliability and Linux source builds. A real shipping cadence, in other words, not a launch-week repo that went quiet.
The bring-your-own-keys model is the sane kind. You plug in your own Claude, OpenAI, Gemini, DeepSeek, Kimi, GLM, Qwen, MiniMax or Doubao keys, or point it at Ollama or LM Studio through an OpenAI-compatible endpoint, and the README says model calls go straight to the provider, never through Orkas servers. There's an optional credits plan if you'd rather not manage keys, and some of those paid tiers are still marked "planned" on the site, along with remote control and team collaboration. So the free open-source core is the shipped product; parts of the paid layer are still promises.
Requirements are ordinary: Node 20+, macOS on Apple Silicon or Intel, Windows 10+, Linux only from source with glibc 2.34+. No musl distros at all.
Two things gave me pause. The packaged desktop app requires an account sign-in on first launch, which is a strange requirement for something calling itself local-first. The source build skips the account layer entirely, so the escape hatch exists if you're willing to build it yourself. And the issue tracker has just two open issues, one of them a Chinese-input bug on WSL that's been sitting since May with no visible maintainer reply. Two issues could mean solid software or it could mean a small user base. With 1.8k stars, I'd guess mostly the latter.
What I have not done yet is run the nine agents through real work and judge the output. That's the only question that actually matters with agent tools, because every one of them demos well. It's a longer job, and it's the follow-up this piece earns if there's interest. What I can tell you today is narrower: every claim I could verify checked out, and the code is open, moving, and buildable without their account or their billing.
If you're already paying for API keys and you're curious what a coordinated multi-agent setup does with them, this is one of the few where you can read the source before you trust it with anything.
Verified September 7, 2026 against the GitHub repo, releases page, issue tracker, and orkas.ai. Re-checked when something moves.
Top comments (0)