Desktop AI commander that dispatches agents and lets them self-improve

What it does

Orkas is a desktop app that puts a “Commander” LLM in charge of specialist sub-agents. You chat with the Commander; it breaks down goals, decides which agents to call, and dispatches them in parallel or in series. Each agent carries its own skills, memory, and task context, and the whole thing—your chats, files, and API keys—stays on your local disk. Model calls go straight from your machine to the provider, never through Orkas servers.

The interesting bit

The clever part is how the agents self-edit. After each task, an agent reflects and updates private meta/COMPETENCE.md and meta/LEARNING_STRATEGIES.md files, then crystallizes useful patterns into reusable skills. On the next run, that metadata is fed back into the system prompt, so the agent literally learns from its own résumé. The Commander also uses “visibility slicing” so agents only see conversation slices relevant to them, saving tokens and preventing private context from leaking across agents.

Key highlights

• Local-first by design: conversations, files, API keys, and agent configs stay on disk; model calls route directly to your provider.
• Bring-your-own-LLM: supports Claude, OpenAI, Gemini, DeepSeek, Kimi, GLM, Qwen, MiniMax, and Doubao, with mixing across agents.
• Structured dispatch via dispatch_to tool calls, deferred wake-up, and a turn-based safety stop (MAX_WORKER_TURNS=100) that caps total worker turns.
• Can drive external CLI coding agents—Claude Code, Codex, OpenCode, Cline—and onboard tools like HyperFrames as local subprocesses.
• Self-evolving agents maintain private meta/ directories and crystallize skills via the skill_manage tool.

Caveats

• Packaged installers are currently only available for macOS and Windows; Linux requires running from source.
• “Fully offline-capable” only applies if you point agents at a local model endpoint; otherwise model calls still need the internet.

Verdict

Worth a look if you want a visual, chat-driven multi-agent setup that keeps data local and improves over time. Skip it if you just need a single chatbot or a code-first framework like LangChain to embed in your own app.