Orkas-AI/Orkas

Orkas is an MIT-licensed desktop application that frames multi-agent orchestration as a chat-driven workspace rather than a code library. You give the Commander a goal; it either does the work itself or routes it to one of 16 built-in specialists — research, content, media, e-commerce, office and coding agents — and to CLI agents you already have installed: Claude Code, Codex, OpenCode, OpenClaw and Hermes-Agent. It suits people who want an ongoing team around a project, not a single chatbot or an embeddable framework.

How the work is coordinated

Routing uses three mechanisms: dispatch_to runs a named agent and returns its result for continued work, hand_off_to transfers the remaining task and ends the Commander's turn, and run_worker runs an anonymous sub-task without conversation history or that agent's skills. Named-agent tasks and anonymous workers have separate concurrency controls, tracked through the dispatch and lock implementations (dispatch and execution locks). Context is budgeted against the selected model's usable input, with a bounded recent-history window and checkpoints that condense long tasks while keeping structured records of file reads, edits and commands (context budgets).

Two features distinguish the workflow from a one-shot assistant. A project carries its goal, rules, memory and files into every new conversation, and with Auto-advance enabled the team picks up the next to-do on its own and returns finished work for review. Scheduled automations can repeat a task hourly through monthly; from v1.9 each automation catches up its most recent missed run within 12 hours when the machine wakes or Orkas starts. Because Orkas runs on your computer, neither feature works while the computer is off.

Install paths and their requirements

The packaged app is the only way to get shared projects, cloud sync and web-app publishing, and it requires signing in — 500 free credits for the built-in model are included (download). It supports macOS 12+ and Windows 10+.

The source build has no account layer and adds Linux, but only glibc 2.34+ x64/arm64; Alpine and other musl distributions are unsupported. It needs Node 22.12+ and a first launch that prepares pinned resources including a ~95 MB embedding model, FFmpeg, whisper.cpp and a speech model. Python 3 and a C/C++ toolchain are needed only when a native package has no prebuilt binary (quick start). Fully offline operation with a local model endpoint is documented for the source build only; the packaged app needs sign-in and cloud models need a network.

Data, permissions and limits

By default data, keys and custom agents stay on disk. With your own provider keys, model calls go directly to the provider; the built-in Orkas model goes through Orkas's managed service, and optional cloud sync in the packaged app stores selected data on Orkas servers (security and data handling). Connectors such as Gmail, Slack and Shopify ask before sending, posting or deleting by default.

Orkas is an orchestrator, not a model. Runtime safeguards — tool-round budgets, repeated-call checks, idle watchdogs and per-task cancellation — contain stalls, but an admitted or queued task is not treated as a completed result. Background reviews of agent activity only run when the app is idle and may decide nothing is worth saving. The author notes it is not intended for those wanting a single chatbot or a code library to embed. The MIT license covers the source; paid plans apply to hosted extras such as credits, sync and app publishing (pricing).