Orkas ships a 16-agent desktop team that runs Claude Code and Codex side by side

The most interesting thing about Orkas, an MIT-licensed desktop app that went up on GitHub as a full multi-agent workspace, is not the agent count. It is the admission buried in the comparison table. Orkas does not claim to beat Claude Code or Codex. It claims to run them. The Commander, Orkas’s planner, hands a coding task to Anthropic’s Claude Code, a review to OpenAI’s Codex or OpenCode, and a release note to its own ContentWriter, all inside one chat. That is a different thesis from most of the agent-startup field. The bet is that orchestration is the scarce layer, not the model.

Orkas is local-first by design. Conversations, files, knowledge bases and custom agents stay on disk by default, and with your own keys the model calls go straight from your machine to the provider. The packaged app adds optional cloud sync, shared projects and web-app publishing. The open-source build does not. That split matters, and we will come back to it.

What actually ships

Sixteen specialists come built in, each with its own skills, memory and tools. They cover research (DeepResearcher, SeoGeoAgent, StockAnalyser), content and media (ContentWriter, PptMaker, ImageStudio, VideoStudio, VoiceStudio, UIDesigner), build and office work (ProductDeveloper, OfficeWorker), and five e-commerce agents (ECommerceResearcher, ECommerceWriter, ECommerceReviewer, ECommerceAnalyzer, ECommerceOperator). A marketplace lists 30-plus more, and the Commander can build a custom agent from a role description.

The coordination layer is the product. Projects hold a goal, rules, memory and files so every new conversation starts from context instead of a blank prompt. A shared to-do list splits work between people and agents. Auto-advance lets the team pick up the next item on its own and bring finished work back for review. Scheduled automations run tasks hourly, daily, weekly or monthly, and from v1.9 each automation catches up its most recent missed run within 12 hours when the machine wakes. Connectors reach Gmail, Google Calendar, Sheets, Drive, Notion, Slack, Shopify and Stripe, and by default Orkas asks before it sends, posts or deletes anything.

That default is the right one. An agent with write access to Stripe and Shopify that acts without confirmation is a liability, not a feature. Orkas’s choice to gate destructive actions behind a prompt is the kind of unglamorous decision that determines whether a tool survives contact with a real business.

The model-agnostic bet

The Commander and all built-in agents share one model setting: the built-in Orkas model, your own provider key, or a local OpenAI-compatible endpoint. The supported key list is long: Claude, OpenAI, Gemini, DeepSeek, Kimi, GLM, Qwen, MiniMax, Doubao. External CLI agents keep their own model settings, so Claude Code stays on Claude and Codex stays on OpenAI’s models while Orkas coordinates around them.

This is a real architectural stance. Most agent products pick a model and build the workflow around it. Orkas treats the model as a swappable component and invests in the planner, the memory and the handoffs instead. If that is right, the value accrues to whoever owns the orchestration and the user’s accumulated project context, not to the lab that owns the weights. If it is wrong, Orkas is a thin wrapper that gets commoditized the moment a lab ships its own multi-agent desktop.

The evidence so far leans toward the first reading, but only partly. Orkas’s own comparison table positions it against LangChain and CrewAI, framing them as code-first frameworks you embed in a Python or JS app while Orkas is a desktop app you direct through chat. That is a fair distinction and a real market: the person who wants a working team of agents without writing orchestration code. It is also a crowded one.

Where the open-source claim gets thin

The MIT license is genuine and the source is readable, which the README stresses (“no black box”). But the features most teams would pay for, shared projects, cloud sync, web-app publishing, live only in the packaged app. The open-source build runs the agents and the Commander but leaves out collaboration. That is a common open-core pattern, and it is defensible. It is also worth naming plainly: the “open-source team of AI agents” headline and the “invite a teammate” feature do not ship in the same artifact.

There is a second gap. Voice input in the source build sends audio to Orkas for recognition and spends the key owner’s Orkas credits. The local-first promise holds for files and conversations, but not for every input path. Read the docs before assuming “local” means “nothing leaves the machine.”

What it means for builders

The interesting signal here is not that another agent app launched. It is that a credible one treats Claude Code, Codex, OpenCode, OpenClaw and Hermes-Agent as interchangeable workers under a single plan. Nous Research’s Hermes-Agent, a self-improving single agent with memory and skills, plugs in as one more CLI session. OpenClaw, a personal-agent platform with messaging integrations and sub-agents, does the same. If the CLI agent becomes the unit of compute and the desktop planner becomes the unit of coordination, the labs compete on being the best worker while someone else owns the job site.

If the CLI agent becomes the unit of compute and the desktop planner the unit of coordination, the labs compete on being the best worker while someone else owns the job site.

That is a plausible near-term structure, and Orkas is one of the first products built explicitly around it rather than around a single vendor’s stack. The risk is equally clear. The labs are already shipping their own desktop and cloud agents, and a coordination layer that depends on their CLIs is exposed to whatever those CLIs change next. Orkas’s answer, per-agent private memory and background reviews that refine self-assessment and learned skills while the app is idle, is an attempt to build switching costs that live in the user’s project context rather than in any one model.

Whether that holds is an empirical question. The 500 free credits and the no-API-key start lower the cost of finding out. The open-core split and the local-first caveats on voice set the ceiling on how much of the promise the free build actually delivers.

Watch the packaged app’s shared projects. If teams start running real client work through a project where a person and an agent share a to-do list, the orchestration thesis gets its first honest test. Until then, Orkas is a well-specified bet on where the value in agents sits, shipped as source you can read.