Orkas is a local-first multi-agent desktop app for macOS, Windows, and Linux. You describe a goal in chat, and a Commander agent plans the work, handles general parts itself, and dispatches specialist agents in parallel or in sequence. A request such as researching competitors and turning the result into a deck runs across DeepResearcher, ContentWriter, and PptMaker and ends with finished files on disk.
Nine specialist agents ship with the app, and more are available from a marketplace or built on request by the Commander:
| Agent | Role |
|---|---|
| DeepResearcher | Plans an investigation, verifies sources, flags contradictions, and writes a cited report |
| ContentWriter | Drafts posts, articles, newsletters, and case studies |
| PptMaker | Builds editable PPTX decks |
| ProductDeveloper | Repository-aware engineering: implements PRDs, fixes bugs, reviews code |
| OfficeWorker | Creates and edits Word, Excel, PowerPoint, and PDF files |
| VideoStudio | Narration, AI presenters, captions, dubbing, and highlight clips |
| ImageStudio | Posters, covers, and infographics, HTML and SVG first |
| UIDesigner | Turns PRDs, screenshots, or Figma material into HTML-first UI |
| SeoGeoAgent | Technical SEO and generative-engine citability audits of a URL |
The Commander is the only actor with dispatch tools. It can hand off to an agent, dispatch and continue, or run a private worker, and a rule forces a hand-off when nothing remains but restating the agent's reply. External CLI agents (Claude Code, Codex, OpenCode, OpenClaw, and Hermes-Agent) can run as local sessions under the same Commander, and open-source projects such as HyperFrames can be onboarded as local tools.
Conversations, files, keys, knowledge bases, and agents stay on your machine, and model calls go straight to the provider. You bring your own keys and can mix providers per agent, including local OpenAI-compatible endpoints such as Ollama or LM Studio. Knowledge bases use a local SQLite index with hybrid full-text and vector retrieval and on-device embeddings. Each agent keeps notes on its own competence and strategies and can turn a solved task into a private skill after a signal-triggered reflection.
Features
- Commander orchestration: plans goals, picks agents, skills, connectors, and tools, and runs specialists in parallel or in sequence
- Built-in specialists: nine agents for research, writing, slides, engineering, office files, video, images, UI design, and SEO audits
- External CLI agents: drives Claude Code, Codex, OpenCode, OpenClaw, and Hermes-Agent as local subprocesses from the same chat
- Per-agent model choice: run one agent on Claude, another on DeepSeek, and another on a local endpoint
- Local knowledge base: SQLite full-text and vector search merged by reciprocal rank fusion, with on-device embeddings
- Model-derived context budget: compaction thresholds scale with the model's context window, with a workspace ledger that survives summarization
- Runaway guards: turn ceilings, per-turn tool-round limits, loop detection, and a tool-idle watchdog
- Self-evolution: agents reflect when weighted signals such as errors or user corrections cross a threshold and can save new private skills
- Local-first storage: conversations, files, keys, and agents stay on disk, and model calls are not proxied
