Orkas is a local-first multi-agent desktop application for macOS, Windows, and Linux. You describe a goal in one chat; a Commander agent plans the work, handles the general parts itself, and coordinates specialist agents in parallel or in sequence, leaving finished files on your disk. Nine specialist agents ship with the app — DeepResearcher, ContentWriter, PptMaker, ProductDeveloper, OfficeWorker, VideoStudio, ImageStudio, UIDesigner, and SeoGeoAgent — out of thirty in a marketplace, and the Commander can build a custom agent from a description.

The Commander is the only actor holding dispatch tools; workers and agents have none. Its three verbs decide who the user hears from and whether the Commander stays in the loop:

Verb	User sees the reply	Result returns to the Commander	Commander’s turn
handoffto	yes	no	ends — the agent’s reply stands as the answer
dispatch_to	yes	yes	continues with a named next step
run_worker	no — private input	yes	continues

Each conversation has a single FIFO runtime and top-level turns run through it serially; parallelism happens inside a turn as nested in-process dispatch, bounded by two semaphores — ten global slots and four nested dispatch slots. Compaction thresholds are derived from the model in use rather than fixed, so one policy covers a 32K window and a 1M one, and a workspace ledger of files written, files read, and commands run survives compaction. Four independent guards bound a runaway run: a 100-turn ceiling, per-turn tool-round budgets, loop detection over near-duplicate tool calls, and a tool-idle watchdog.

A knowledge base is a SQLite index built on the machine, with FTS5 for full text and stored vectors for similarity. Documents are chunked at 512 lines with 64 lines of overlap, and a query runs both ways with the ranked lists merged by reciprocal rank fusion weighted 0.7 vector to 0.3 keyword. Embeddings are computed on device by a bundled ONNX model, with cloud embedding providers available instead. Each agent also keeps notes on its own competence and working methods, reviews itself when a weighted score over six signals clears a threshold, and can crystallize a solved problem into a private skill.

Beyond its own agents, Orkas drives external command-line coding agents as local subprocesses and onboards open-source projects as local tools, all from the same chat. Model keys are your own — Claude, OpenAI, Gemini, DeepSeek, Kimi, GLM, Qwen, MiniMax, and Doubao, or a local endpoint added as an OpenAI-compatible custom provider — and different agents in one chat can run on different providers. Conversations, files, API keys, knowledge bases, and custom agents stay on disk, and model calls go straight from the machine to the provider.

Features

Commander dispatch: one planning agent breaks a goal into steps and routes them to specialists concurrently or in sequence
Nine built-in agents: research, writing, slides, engineering, office documents, video, images, UI design, and SEO auditing, ready on first launch
Agent marketplace: thirty agents to install, or a custom agent the Commander builds from your description
External CLI agents: command-line coding agents run as local subprocesses under the same chat
Bring-your-own models: Claude, OpenAI, Gemini, DeepSeek, Kimi, GLM, Qwen, MiniMax, and Doubao, or a local OpenAI-compatible endpoint, mixed per agent
Local-first storage: conversations, files, keys, knowledge bases, and agents stay on disk; model calls are never proxied
Hybrid memory: a local SQLite knowledge base combining FTS5 and vector search by reciprocal rank fusion, reached through memorysearch and memoryread
On-device embeddings: a bundled ONNX model indexes a knowledge base without sending anything out, with cloud providers optional