Orkas

A local-first desktop workforce where a Commander coordinates specialist agents for complex work.

Source repo
Orkas-AI/Orkas
Stars
★ 2.1k
Last updated
1d ago
License
MIT
Primary language
TypeScript

At a glance

Works with
Universal · cross-platformCodex · Claude Code · OpenAI API · Claude API
You'll need
Node 20+Python 3macOS, Windows 10+, or recent LinuxShell / CLINetwork accessLocal filesystem
Typical use
A product manager wants to turn a PRD into actionable development tasks and distribute the work among specialists.
Main limitation
Model calls still need network access unless a local model endpoint is configured; offline-capable does not make cloud models available offline.

What does this agent do, and when should you use it?

Orkas is an MIT-licensed, local-first desktop application for people who need coordinated agent work on complex tasks. In one chat, its Commander interprets a goal, breaks it down, and dispatches specialist agents when needed, while also handling general analysis, writing, research, file work, and automation itself. Agents can work in parallel or in sequence with focused skills, memory, and task context, while collaboration progress is shared through plan.md. The application runs on macOS, Windows, and Linux; conversations, files, knowledge bases, agent configurations, and model keys remain on the local machine, while model calls go directly to the chosen provider. It supports providers including Claude, OpenAI, Gemini, DeepSeek, Kimi, GLM, Qwen, MiniMax, and Doubao, and can drive local CLI coding agents such as Claude Code, Codex, OpenCode, and Cline.

A user submits a goal in the main chat, and the Commander uses the relevant context before staging work for specialists with dispatch_to({to, message}); recipients wake after the Commander’s current turn finishes. Every dispatch is routed through enqueue, and the Commander records collaborative progress in the shared plan.md. The main conversation is stored as jsonl, while workers see only messages matching from==me, a recipient list containing them, or mentions containing them. Each agent keeps meta/COMPETENCE.md and meta/LEARNING_STRATEGIES.md in its own directory and can use skill_manage to crystallize a solution into a private skill. Orkas can also run Claude Code, Codex, OpenCode, and Cline as local subprocesses.

  1. A product manager wants to turn a PRD into actionable development tasks and distribute the work among specialists.
  2. A researcher needs to gather, summarize, and deliver recurring market research while keeping working files on the local machine.
  3. A team handling sensitive local material wants to drop files into a desktop app for document chat or data analysis.
  4. A developer wants to coordinate Codex, Claude Code, OpenCode, or Cline from the same conversational workflow.
  5. A small team needs one coordinated workflow that can produce code, research, video, and slide decks using local tools.

How do you install or deploy this agent?

Packaged installers are available from orkas.ai for macOS and Windows; Linux currently requires running from source. Source execution requires Node 20+, Python 3, and macOS, Windows 10+, or recent Linux:

git clone https://github.com/Orkas-AI/Orkas.git
cd Orkas
./run.sh

On Windows, use run.cmd. The scripts install dependencies and download an approximately 95 MB embedding model. After first launch, add a model-provider API key or OAuth in Settings → AI Providers.

How do you use this agent?

After first launch, configure an API key or OAuth in Settings → AI Providers, then describe the goal in the main chat. The Commander either completes general work directly or dispatches specialist agents as needed, with shared progress recorded in plan.md. Claude Code, Codex, OpenCode, and Cline can be used as CLI coding agents in the same workflow.

What are this agent's strengths and limitations?

Pros
  • Local-first storage keeps conversations, files, knowledge bases, agent configuration, and keys on the machine; model calls are not proxied or archived by Orkas servers.
  • It provides Commander-led multi-agent coordination through a desktop chat instead of requiring orchestration code.
  • It supports multiple model providers and can mix providers across agents.
  • It can incorporate Claude Code, Codex, OpenCode, and Cline as local subprocesses from the same chat workflow.
  • Per-agent meta/ memory and private skills support reflection and reuse after tasks.
Limitations
  • Model calls still need network access unless a local model endpoint is configured; offline-capable does not make cloud models available offline.
  • Users must supply their own provider API keys or OAuth and pay their model providers.
  • Linux has no packaged installer and currently requires running from source.
  • Source startup requires Node 20+ and Python 3, and first launch downloads an approximately 95 MB embedding model.
  • The supplied documentation does not establish support for enterprise identity, centralized key management, or cloud-based team sharing.

How does this agent compare with similar options?

Compared with code-first frameworks such as LangChain and CrewAI, Orkas offers a local desktop chat interface for coordinating work. Compared with cloud agent orchestrators, it keeps data and keys on the local machine. It is distinct from a single always-on assistant such as OpenClaw or a single-assistant experience such as Claude Desktop: Orkas coordinates multiple specialists through its Commander, while OpenClaw and Hermes-Agent can serve as CLI backends.

Key facts side by side with the most closely related agents.

Agent Source review Stars Updated Language Full support on
Orkas This agent 43 · Major gaps ★ 2.1k 1d ago TypeScript Codex · Claude Code · OpenAI API · Claude API
NextClaw 58 · Major gaps ★ 259 4d ago TypeScript Codex · Claude Code
MyAgents 57 · Major gaps ★ 896 5d ago TypeScript Codex · Claude Code · Claude API
BagIdea Office 71 · Some gaps ★ 230 1d ago JavaScript Claude Code · OpenAI API

How does FollowAgents rate this agent?

FollowAgents source review · FARS-2.1
Major gaps
43/ 100 5-point scale 2.2 / 5
Trust 8/29
Reliability 6/14
Adaptability 10/18
Convention 9/18
Effectiveness 7/13
Verifiability 3/8
Why each dimension lost points
Trust8 / 29 · 1.4/5

Evidence shows: README claims local-first, data and keys stay local, model calls go directly to provider; but no implementation details on permission management, user confirmation, or data flow transparency. Deductions: no user confirmation mechanism, insufficient sensitive data handling details, no dependency security audit, external effects (e.g., network requests) not clarified, no rollback mechanism, source attribution only mentions acknowledgments to OpenClaw and Hermes-Agent without detail.

Reliability6 / 14 · 2.1/5

Evidence shows: README and code structure are consistent, description matches functionality; but dependency availability not verified (e.g., runtime downloads), failure messages not documented. Deductions: dependency availability relies on auto-install scripts without offline or fallback options; failure messages not documented.

Adaptability10 / 18 · 2.8/5

Evidence shows: README clearly identifies target users (researchers, data analysts, developers) and use cases; capability boundaries described (not a local LLM, not a code framework); environment fit (macOS/Windows/Linux) stated. Deductions: trigger precision (e.g., distinction between @ and dispatch_to) only mentioned in README without detailed docs; environment fit only lists system requirements without detailed configuration guide.

Convention9 / 18 · 2.5/5

Evidence shows: README structure clear, includes quick start, FAQ, comparisons; install notes detailed (download and source run); naming stable (Orkas); examples and FAQ rich; known limitations mentioned (Linux requires source run); MIT license explicit; version in package.json (2026.7.29) but no CHANGELOG; maintenance responsibility not clear (no contribution guide or maintainer). Deductions: no CHANGELOG, maintenance responsibility unclear.

Effectiveness7 / 13 · 2.7/5

Evidence shows: output usability (desktop app, GUI) and marginal value (multi-agent coordination) emphasized in README; cost-benefit (BYO LLM keys) stated. Deductions: cost-benefit lacks concrete data (e.g., performance benchmarks, resource usage), marginal value not quantified compared to existing tools.

Verifiability3 / 8 · 1.9/5

Evidence shows: README claims (e.g., local-first, multi-agent) have some code support (e.g., visibility slicing), but no test results or independent verification; cross-source corroboration insufficient (only internal tests); facts and inferences not clearly separated. Deductions: claims lack traceability, tests only exist for some skill modules, not covering core functionality.

Risks and how to mitigate them
  • Not found in source: confirmation before actingTurn on (or add) a confirmation step before it acts, and try it in a sandbox or test environment before real data.
  • Not found in source: rollback or recovery pathBack up first, or work on a git branch or snapshot, so its changes can be undone.
  • Publisher identity unverified, proceed with caution.
  • Dependency security not audited, supply chain risk.
  • User confirmation mechanism missing, tool calls may occur without user consent.
  • Sensitive data handling details insufficient, review key storage.
  • Rollback mechanism not mentioned, upgrade failure may not be recoverable.
Evidence confidence: Low Reviewed Aug 09, 2026 Reviewed revision 5a86844a9152 New commits since this review; the score may not cover them
See the full review method →

FAQ

Is Orkas a local LLM?
No. It runs locally but calls models through the API keys you configure, or through a local model endpoint.
What does it cost?
Orkas is MIT licensed; you pay the model providers you choose to use.
Does Orkas send my data to its servers?
The README states that conversations, files, knowledge bases, agents, and keys remain local, and model calls go directly to the selected provider rather than through or into Orkas servers.
Can it work fully offline?
The app is offline-capable, but model calls need network access unless agents are pointed at a local model endpoint.
View on GitHub ↗ Install ↓

Compare agents like this one

The same FARS review applied across the shortlist this agent qualifies for.

Related agents