Orkas: A desktop multi-agent team that turns goals into ongoing work
Orkas is a desktop multi-agent workspace that turns goals into ongoing to-dos instead of answering one prompt.
GitHub Orkas-AI/Orkas Updated 2026-11-01 Branch main Stars 6.2K Forks 376
JavaScript/TypeScript Desktop multi-agent Commander macOS/Windows/Linux

🧭 Decision Guide

Why trending now: The README combines a local-first desktop multi-agent app, Commander, 16 built-in agents, Claude Code/Codex/OpenCode coordination, Connectors, and 500 credits; the repository gained 1,044 stars that day and lists Orkas 2026.10.27 as its latest release, but the material does not prove which factor directly caused the daily trend.

Try it if you

  • You want Commander to manage development work involving Claude Code, Codex, and OpenCode.
    The README section “Run Claude Code, Codex, OpenCode, OpenClaw and Hermes-Agent together” says these CLI agents can run as coordinated local sessions.
  • Your work needs Projects, Auto-advance, and recurring hourly-to-monthly automations.
    The README section “Keep the work moving” lists project memory, automatic to-do advancement, and scheduled automations.
  • You want conversations, files, and knowledge bases to remain local by default while using your own model keys.
    “Local-first by design” in “Why Orkas” says data stays on disk and calls using your own keys go directly from the machine to the provider.

Skip it if you

  • You need to run the source build on Linux musl, Alpine, or glibc below 2.34.
    Quick start explicitly requires glibc 2.34+ on Linux and states that Alpine and other musl distributions are unsupported.
  • You need cloud sync, Shared projects, or web-app publishing in the open-source build.
    The README NOTE states that these features are only in the packaged app and are not supported by the open-source build.
  • Your data policy cannot allow voice audio to be sent to Orkas.
    The README Voice input section says audio is sent to Orkas for recognition and uses the key owner's Orkas credits.

Requirements

  • Node 22.12+
  • macOS / Windows 10+ / glibc 2.34+ Linux x64 or arm64
  • Python 3 and a C/C++ build toolchain for the uncommon native-module source-build fallback
  • The source build requires your own keys or a local endpoint for models
  • Linux source runs do not support Alpine or other musl-based distributions

First step (verbatim from README)

git clone https://github.com/Orkas-AI/Orkas.git

Watch out

  • The first source launch prepares pinned Python, uv, Node, FFmpeg, whisper.cpp, and a roughly 95 MB embedding model.
    The README Quick start section's run.sh / run.cmd description.
  • For native dependency errors, the README requires Node.js 24 LTS before running npm run native:repair.
    The Native dependency errors paragraph in README Quick start.
  • After wake or startup, an automation catches up only its most recent missed run within a 12-hour window.
    The Scheduled automations bullet in the README, with the v1.9 note.
  • Connectors ask for confirmation before sending, posting, or deleting by default.
    The Connectors bullet in the README.

Alternatives

  • Use Claude Code directly：Better when you only need one coding agent and do not need Commander, Projects, or 16 built-in agents.
    README section “Run Claude Code, Codex, OpenCode, OpenClaw and Hermes-Agent together”
  • Use Codex directly：Better when you want an independent Codex coding flow without Orkas's cross-agent plan and to-dos.
    README section “Run Claude Code, Codex, OpenCode, OpenClaw and Hermes-Agent together”

Not stated in the README

  • The README does not provide the complete names and individual capabilities of the 16 built-in specialist agents.
  • The README does not provide performance data for Commander across models, task sizes, or concurrency levels.
  • The README does not specify the exact built-in Orkas model, context window, or pricing rules.
  • The README does not document the full permission scope or OAuth failure handling for Gmail, Slack, Stripe, and other connectors.
  • The README does not provide production cases or reliability metrics corresponding to 6,217 stars, 10 contributors, and 5 releases.
  • The README does not state whether features and performance are identical across Windows, macOS, and Linux.

💡 Deep Analysis

6
Yes I can deploy only on glibc 2.34+ Linux x64/arm64, not Alpine, and I can provide Node 22.12+, Python 3, and a C/C++ toolchain. Is Orkas’s source-based setup suitable for me?
For: A developer running from source on glibc 2.34+ Linux x64 or arm64 who cannot use the packaged macOS or Windows app

Yes, provided you accept that the first launch prepares many local dependencies rather than behaving like a lightweight Linux binary package.

  • The README lists source execution as the current way to run Orkas on Linux and supports glibc 2.34+ Linux x64/arm64.
  • It requires Node 22.12+ and recommends Python 3 plus a C/C++ toolchain; first launch also prepares Python, uv, an embedding model, OfficeCLI, FFmpeg, whisper.cpp, and speech-model resources.
  • Alpine and other musl distributions are explicitly unsupported.
  • If Electron native modules fail, the README provides npm run native:repair; Linux is supported, but native ABI and dependency maintenance remain deployment requirements.
  • Download: Run from source: macOS, Windows, glibc 2.34+ Linux x64/arm64
  • Quick start: Node 22.12+ · macOS / Windows 10+ / glibc 2.34+ Linux
  • “Alpine and other musl-based distributions are not supported.”
  • Quick start: `npm run native:repair`
./run.sh           # macOS / Linux
Not stated in the README：The README does not specify supported Linux distributions, minimum disk space, total download size, or first-launch duration.；The README does not explain operation on headless systems or remote servers without a desktop environment.
Yes I use macOS 12+ and do not want to configure a third-party API key first. I want to install the desktop app and start managing projects and to-dos with the built-in Orkas model. Is this path suitable?
For: An individual developer on macOS 12+ with Apple Silicon or Intel who wants the free packaged desktop app and built-in model

Yes, because macOS 12+ has a packaged build, and starting with the built-in model does not require a third-party API key first.

  • The README’s Download table lists macOS 12+ with both Apple Silicon and Intel support.
  • After signing in, the packaged app provides 500 free credits for the built-in Orkas model.
  • Projects retain goals, rules, memory, and files; To-dos divide work between people and agents, supporting ongoing projects rather than one-off chats.
  • Note that the packaged app requires sign-in and the built-in model runs through Orkas’s managed model service. If you require no account or fully offline operation, this path is unsuitable.
  • Download: macOS 12+ (Apple Silicon & Intel)
  • README: free, with 500 credits for the built-in Orkas model
  • Download: Packaged app (recommended), Account: Sign in — 500 free credits
  • README: A project holds the goal, rules, memory and files
Not stated in the README：The README does not explain how many calls 500 credits represent, their expiration, or task-specific pricing.；The README does not explain packaged-app behavior when offline, signed out, or out of credits.
Yes I already use Claude Code, Codex, and OpenCode locally. I want one project plan to assign bug fixing, code review, and release-note drafting to them. Is Orkas suitable for replacing my manual CLI switching?
For: A software developer who already has Claude Code, Codex, and OpenCode installed and wants one plan to coordinate bug fixes, reviews, and release notes

Yes, because Orkas’s Commander can turn one goal into a plan and assign steps to local CLI agents and built-in specialists.

  • The README explicitly supports Claude Code, Codex, OpenCode, OpenClaw, and Hermes-Agent as installed local CLI sessions.
  • The Commander “hands each one the part it’s best at and keeps one plan across all of them,” so fixing, review, and release-note work can share one plan.
  • Projects retain goals, rules, memory, and files so later conversations inherit the context.
  • However, external CLIs keep their own model settings, login state, and environment; Orkas does not remove dependency or handoff risks.
  • Run Claude Code, Codex, OpenCode, OpenClaw and Hermes-Agent together
  • “keeps one plan across all of them”
  • Why Orkas: “A Commander that plans, delegates and does the rest”
  • Project insight: scheduler and execution agents are decoupled
./run.sh           # macOS / Linux
Not stated in the README：The README does not explain how structured output is passed between multiple CLI agents or whether conflicting results are resolved automatically.；The README provides no performance, cost, or success-rate data for different CLI combinations.
Yes I have a collection of local documents and need a research project to persist project rules, agent memory, document retrieval, and chat history instead of repeating the background every time. Can Orkas meet this context-management requirement?
For: An individual researcher who wants local documents, project memory, and chat history to support long-running research or content projects while separating global and project-specific context

Yes, because Orkas provides persistent context layers for project state, memory, document indexing, and conversation history.

  • The README says a project stores its goal, rules, memory, and files, so new conversations pick up the previous context.
  • Durable notes separately store user preferences, project facts, and reusable agent knowledge; these are distinct from the Library index and conversation history.
  • The Library extracts document passages, builds a local SQLite vector index with a local embedding model, and can search relevant passages and read source content.
  • Retrieval quality still depends on extraction, chunking, and embedding quality; the README does not guarantee recall accuracy for scans, complex tables, or large collections of duplicate documents.
  • Orkas introduction: A project holds the goal, rules, memory and files
  • How it works / Memory and the Library
  • “Its embedding model runs on your machine”
  • “stores a local SQLite vector index”
Not stated in the README：The README does not specify document-count, file-size, or vector-index capacity limits per project or Library.；The README does not specify OCR quality for scans, table-parsing capabilities, or retrieval evaluation metrics.
It depends I want conversations, files, knowledge bases, and custom agents to stay on my machine by default, while using a local OpenAI-compatible endpoint for inference. Can Orkas meet this data-control constraint?
For: A privacy-sensitive developer who wants data kept locally and prefers a local OpenAI-compatible endpoint over a hosted model

It depends: local storage and local endpoints are both documented, but that does not mean every feature works fully offline.

  • The README says its local-first design keeps conversations, files, knowledge bases, and custom agents on disk by default.
  • The model layer supports a local endpoint; with your own keys, requests go directly from the machine to the provider.
  • The Library runs its embedding model locally and stores document retrieval data in a local SQLite vector index.
  • However, the packaged Orkas model uses Orkas’s managed model service, while speech recognition, connectors, OAuth, cloud sync, and public publishing may require network access or accounts. Local storage and offline operation must therefore be treated separately.
  • Why Orkas: “Local-first by design”
  • Download: Built-in models or your own keys; “a local OpenAI-compatible endpoint”
  • How it works: Library stores “a local SQLite vector index”
  • Project insight: local-first does not mean every capability is offline
git clone https://github.com/Orkas-AI/Orkas.git
Not stated in the README：The README does not enumerate which connectors or features send data to remote services, nor the authentication and TLS details for local endpoints.；The README does not provide a complete offline-mode or network-blocking feature matrix.
It depends I want agents to handle Gmail, Google Calendar, Notion, Slack, Shopify, or Stripe tasks on a schedule, while requiring confirmation before sending, publishing, or deleting. Is Orkas suitable for this automation?
For: A small-business automation user handling Gmail, Google Calendar, Sheets, Drive, Notion, Slack, Shopify, or Stripe tasks hourly, daily, weekly, or monthly

It depends: Orkas covers scheduling, connectors, and confirmation for risky actions, but the README does not establish it as a business-process engine safe to run without supervision.

  • The project supports Gmail, Google Calendar, Sheets, Drive, Notion, Slack, Shopify, and Stripe connectors.
  • Scheduled automations can run hourly, daily, weekly, or monthly, and missed runs may be caught up after restart or wake-up within a limited window.
  • The README and project insight say sending, publishing, and deleting request confirmation by default, matching your approval-gate requirement.
  • However, Auto-advance and scheduled tasks may continue with unintended follow-up steps; connector reliability, authorization scope, failure alerts, and production guarantees are not established by the README.
  • Project insight: connectors interact with external business systems
  • Key features: Gmail, Google Calendar, Sheets, Drive, Notion, Slack, Shopify, and Stripe
  • README: scheduled automations repeat routine work hourly, daily, weekly or monthly
  • Project insight: confirmation is requested before sending, publishing, or deleting
Not stated in the README：The README does not specify the exact operations, OAuth scopes, retry policies, or rate limits for each connector.；The README does not specify failure notifications, audit-log formats, or idempotency guarantees for scheduled tasks.

✨ Highlights

  • Commander coordinates 16 built-in specialist agents
  • Supports Claude Code, Codex, and OpenCode
  • Projects retain goals, rules, memory, and files
  • Automations support hourly, daily, weekly, and monthly schedules
  • The open-source build lacks cloud sync and Shared projects

🔧 Engineering

  • Commander breaks goals into to-dos and routes them to specialist agents.
  • Projects carry rules, memory, and files into every conversation.
  • Auto-advance takes the next to-do and returns completed work for review.
  • Connectors cover Gmail, Slack, Notion, Shopify, and Stripe.
  • Local-first storage keeps conversations, files, libraries, and custom agents on disk.

⚠️ Risks

  • Linux source runs require glibc 2.34+ and do not support musl.
  • The open-source build lacks cloud sync, shared projects, and web publishing.
  • The built-in Orkas model routes calls through Orkas's managed model service.
  • First launch downloads a roughly 95 MB embedding model and other resources.
  • Voice audio is sent to Orkas and consumes credits.

👥 For who?

  • Teams that need continuous project work on macOS or Windows desktops.
  • Developers already using Claude Code, Codex, or OpenCode.
  • Individuals connecting workflows through Gmail, Slack, or Notion.
  • Source users who can provide Node 22.12+ and Python 3.