Model or dataset
Orkas-AI/Orkas avatar
Orkas-AI/Orkas

Orkas: a Commander LLM, 16 specialist agents, and the limits of a desktop that has to stay awake

Open-source multi-agent AI desktop client — build and command your AI agent team through conversation. A commander LLM dispatches sub-agents in parallel or in series; agents self-evolve via reflection and skill crystallization. Local-first, BYO LLM keys (Claude · OpenAI · Gemini · DeepSeek · Kimi · GLM · Qwen). macOS / Windows / Linux.

9,377 stars555 forksTypeScriptMIT

At a glance

What is it?
Orkas is an MIT-licensed Electron and Node desktop app in which a Commander model plans work and hands parts of it to specialist agents and to coding agents such as Claude Code and Codex. It ships often, but its automations, model credits and platform downloads each carry conditions a buyer should know first.
Who is it for?
Orkas suits people who want a local desktop app that coordinates many task-specific agents across ongoing projects, and who can keep a computer running for scheduled work. It is a poor fit for Linux users who want a packaged installer, for anyone expecting automations to run while the machine is off, and for video work that depends on pulling clips from social links.
Can I use it commercially?
Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Yes. The repository received new commits within the last day.
What is it written in?
Mainly TypeScript, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • Orkas-AI/Orkas9,377 ★ · +7,497
Star history
Repository2026-09-102026-11-15
Orkas-AI/Orkas1.9k9.4k

7,497 Stars in the last 66 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

The Commander plans the work and pulls in outside coding agents

Every request in Orkas goes to the Commander. Small tasks it finishes itself; larger ones it splits into steps, assigns each part to a specialist agent, runs those agents in parallel or in series, and combines the results. The example given for this is market research that ends up as a report, a slide deck and a launch video. The Commander can also bring in coding and personal agents that live outside Orkas: Claude Code, Codex, OpenCode, OpenClaw and Hermes-Agent. Those are agents you already use, so Orkas coordinates them rather than bundling them. If one of them is not installed and configured on your machine, the Commander has nothing to hand that part of the job to. Agents are also described as improving themselves through reflection and skill crystallization, and when no built-in agent fits, you can describe a role and the Commander builds a custom agent.

Sixteen built-in specialists, each with its own boundary

The app ships with 16 agents, each carrying its own skills, memory and tools, and an agent marketplace lists more than 30. The research group is DeepResearcher, SeoGeoAgent, which takes a URL and returns a technical audit, Core Web Vitals and a ranked fix list, and StockAnalyser for equity research, factor screens and backtests. Content and media covers ContentWriter, PptMaker for editable PPTX decks, ImageStudio, VideoStudio, VoiceStudio and UIDesigner, which produces HTML-first UI from PRDs, screenshots or Figma material. ProductDeveloper does repo-aware engineering and OfficeWorker handles Word, Excel, PowerPoint and PDF files. Five agents cover e-commerce. The boundaries matter: ECommerceOperator plans catalog, price, promotion and inventory changes but executes only through tools you connect, and ECommerceAnalyzer works from your exports. Without connected tools or exported data, those two agents produce plans and analysis, not changes to your store.

VideoStudio cannot fetch a reference video from a share link

VideoStudio can rebuild a reference video's narrative structure, visual style and pacing around your own product, script and footage. It analyzes the hook, information order, visual beats and call to action, then proposes what to keep and what to replace before producing anything. The constraints are stated plainly. You must start with a video file the agent can read, because downloading from social-platform share links is not built in. Exact camera moves, human performances and frame-by-frame matching are not guaranteed. Locally authored text, captions, narration and clips can be revised on their own, but text baked into AI-generated footage can only be changed by changing that footage. The consequence is that this workflow fits motion-graphic explainers, app demos and product ads built from your own material, not shot-for-shot copies of live-action content.

Projects carry the context so conversations do not start cold

A project in Orkas holds the goal, rules, memory and files, and every new conversation inside it starts from that state, so you do not brief the agents from scratch each time. Each project also has a to-do list shared between people and agents. With Auto-advance switched on, the team picks up the next to-do on its own and returns finished work for your review. That review step is part of the design: work comes back to you rather than being published or sent onward automatically, which keeps a human approval point in place even when the queue advances without you.

Scheduled automations only run while your computer does

Automations can repeat any task hourly, daily, weekly or monthly, but Orkas runs on your computer, not on a hosted scheduler. From v1.9, each automation catches up only its most recent missed run, and only within 12 hours, when the computer wakes or Orkas starts. A laptop that is closed over a weekend will therefore not replay every hourly run it missed, and a run missed by more than 12 hours is not caught up at all. Anyone planning reports or monitoring on a schedule needs a machine that stays on, or has to accept gaps. Note too that the current releases use date-based tags such as v2026.10.27, while the catch-up behavior is pinned to v1.9, and the mapping between the two numbering schemes is not spelled out next to that feature.

Starter credits, then your own keys or a local endpoint

The download is free and includes 500 credits for the built-in Orkas model, so no API key is needed to begin. After that, or from the start if you prefer, you bring your own keys for Claude, OpenAI, Gemini, DeepSeek, Kimi, GLM, Qwen, MiniMax or Doubao, or point Orkas at a local OpenAI-compatible endpoint. The project calls itself local-first and says all your data stays on your computer. Which provider you pick decides where prompts are processed, so the local endpoint option is the one to evaluate if keeping prompts off third-party services matters to you.

The installer button covers macOS and Windows, not Linux

The repository description lists macOS, Windows and Linux, but the download link on the project page is labeled for macOS and Windows. Linux users should expect to work from the source tree. The package.json names the app as an Electron and Node one-piece app with bootstrap.cjs as its entry point and declares [email protected] as its package manager. The start and test scripts include:

code
"start": "electron .",
"test:non-model": "node scripts/run-non-model.mjs",

The prestart hook runs scripts/ensure-dev-dependencies.cjs first. The stop script is the catch: it calls pkill against a path ending in Orkas.app/Contents/MacOS/Electron, a macOS application bundle layout, and ends in || true. On Linux or Windows that pattern will not match the running process and the command still exits successfully, so npm run stop gives no sign that nothing was stopped.

Date-stamped releases arrive every few weeks

The repository is not archived and its last push was on 1 November 2026. Recent releases are v2026.10.1 on 1 October, v2026.10.17 on 17 October and v2026.10.27 on 27 October 2026, and package.json carries the matching version 2026.10.27. The source tree includes test/, eval/ and three Playwright configurations, one of them a release harness, plus a test:resources script that runs Python tests over resources/builtin. Contributors building from source therefore need a Python setup as well as Node to run the full test suite. Documentation is available in English, Simplified Chinese and Japanese README files.

Editorial conclusion

Orkas suits people who want a local desktop app that coordinates many task-specific agents across ongoing projects, and who can keep a computer running for scheduled work. It is a poor fit for Linux users who want a packaged installer, for anyone expecting automations to run while the machine is off, and for video work that depends on pulling clips from social links. Before relying on it, decide which model provider or local OpenAI-compatible endpoint you will use after the 500 starter credits, and check that the outside coding agents you want the Commander to call are already installed.

Frequently asked questions

Which AI models can Orkas use?

Orkas includes 500 credits for its built-in model, and it accepts your own keys for Claude, OpenAI, Gemini, DeepSeek, Kimi, GLM, Qwen, MiniMax and Doubao, or a local OpenAI-compatible endpoint.

Do Orkas scheduled automations run when my computer is off?

No. Orkas runs on your computer, and from v1.9 each automation catches up only its most recent missed run within 12 hours when the computer wakes or Orkas starts.

Can Orkas VideoStudio download a reference video from a social media link?

No. Social-platform share-link downloading is not built in, so you need to supply a video file the agent can read.

Is Orkas spelled with a k?

Yes. The project is Orkas, published by the Orkas-AI organization on GitHub with its homepage at orkas.ai.

Official sources

  1. License: MIT
  2. Orkas-AI/Orkas on GitHub
  3. Project website
  4. README
  5. Releases
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/orkas-ai-orkas.svg)](https://hysenlabs.com/projects/orkas-ai-orkas)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/orkas-ai-orkas.svg)](https://hysenlabs.com/projects/orkas-ai-orkas)

Was this useful?