Orkas Orkas
Download GitHub
Home Docs

Orkas Documentation Orkas 文档 Orkas ドキュメント

Everything you need to install Orkas, connect your own model, and direct an AI team — plus how the local-first, multi-agent architecture works under the hood. 安装 Orkas、接入你自己的模型、指挥一支 AI 团队所需的一切 —— 以及本地优先的多 agent 架构在底层是如何运作的。 Orkas のインストール、自分のモデルの接続、AI チームの指揮に必要なすべて。そして、ローカルファーストなマルチエージェント・アーキテクチャが内部でどう動くかを解説します。

# Introduction

Orkas is an open-source, local-first desktop AI client. Instead of chatting with a single assistant, you direct a team: a lead agent owns your goal and recruits sub-agents, each of which calls skills to get work done. You describe the goal in chat instead of drawing a flowchart or setting up automation rules.

It runs on macOS and Windows. Your model keys stay on your machine, and your workspace is local-first by default; you bring your own LLM provider keys.

New here? Read this page top to bottom for a complete mental model, or jump straight to Install to get running in two minutes.

# Install

Download the desktop client and open it — no account required to start.

1

Download

Get the build for your OS from the download page or GitHub releases.

2

Open

Launch Orkas. On first run it sets up a local data directory on your machine.

3

Stay current

The app checks for updates; the open-source build can also be pulled and run from source.

# run from source (open-source build)
git clone https://github.com/Orkas-AI/Orkas
cd Orkas/PC && ./run.sh

# Connect a model

Orkas is bring-your-own-model. Open Settings → Models and add a provider via OAuth or an API key. Traffic goes directly from your machine to the provider — it is never proxied through Orkas servers.

ProviderHow you connectWhere your data goes
OpenAI (GPT)OAuth or API keyDirect to provider · never proxied
Anthropic (Claude)OAuth or API keyDirect to provider · never proxied
Google (Gemini)OAuth or API keyDirect to provider · never proxied
MiniMaxAPI keyDirect to provider · never proxied
Kimi (Moonshot)API keyDirect to provider · never proxied
Zhipu (GLM)API keyDirect to provider · never proxied
+ more, added over timeOAuth or API key100% local — keys on your device
Security: keys are encrypted at rest (AES-256-GCM) in a local config that never syncs. See Security & Trust for the full key lifecycle.

# Your first task

Type a goal in the Commander chat. The lead agent plans, recruits the sub-agents it needs, and streams results back. Risky steps pause for your approval.

you › turn this product idea into a launch checklist and a first-week plan
› lead agent recruiting: planner, writer
planner › breaks work into milestones and owners
writer › drafts announcement copy and open questions
✓ plan ready — review it in chat

# Architecture

One task flows through three layers. Each actor has a clear responsibility and a bounded context, so a single request can span disciplines without overflowing one context window.

You

Direct

State a goal in natural language.

Lead agent

Orchestrate

Owns the full context; recruits sub-agents and splits the work.

Sub-agents

Execute

Each works in its own context, calling skills to deliver.

Why a team, not one big thread: bounded sub-agent contexts keep token cost down and responsibilities clean — the lead passes each sub-agent only what it needs.

# Lead & sub-agents

The lead agent (Commander) holds the global thread and decides who does what. Sub-agents are specialists you can also summon directly with @name. Turn a recurring task into a sub-agent once, then summon it forever — that's "sub-agents as workflows".

  • Auto-assembly — the lead picks sub-agents by task and capability.
  • Context boundaries — sub-agents see only what the lead passes plus your direct messages.
  • Self-evolution — via meta-cognition, agents review their runs and improve next time (on by default).

# External agents

Already living in a coding CLI? Orkas can drive external CLI coding agents as local subprocesses and manage them alongside its own native team — so you summon them in chat like any other agent and direct everything from one interface.

CLINotes
Claude CodeAnthropic's coding CLI.
CodexOpenAI's coding CLI.
OpenClawOpen-source coding CLI.
HermesCoding CLI.
OpenCodeOpen-source coding CLI.
  • Install it yourself — you install the CLI; Orkas does not bundle or install it for you.
  • Auto-detect — Orkas scans your PATH for the binary (or a custom path you set via ORKAS_<TYPE>_PATH), runs --version, and checks a minimum version.
  • Attach & run — pick a detected CLI as an agent's runtime; Orkas spawns it as a local subprocess inside your conversation, managed next to native agents.
Stays local: the CLI runs on your machine and your keys and data stay local. See Security & Trust.

# Skills

Skills are an agent's arms — capabilities like web search, code execution, file I/O, image generation, and connector calls. Use the built-ins, write your own, or install from the Marketplace.

SkillWhat it does
web_searchQuery the web through a search provider, return structured results.
code_execRun shell / code in a sandbox to reproduce, build, or test.
file_ioRead and write files in the working directory.
kb_searchSemantic + keyword search over your local knowledge base.
create_artifactGenerate a self-contained interactive app, run it sandboxed.

# Knowledge base

Drop PDFs, markdown, and docs into a local knowledge base. Files are chunked and indexed on your machine; agents kb_search them to answer with citations. The source files stay local — only your model call leaves the device.

# Memory

Orkas keeps personal memory across sessions in two local files: preferences (who you are, how you like things) and facts (decisions, project context). Both are editable, importable, and injected at session start so the team compounds knowledge about how you work.

# Tech stack

Orkas is an Electron desktop app. The lead and sub-agent runtime runs inside the app, and your LLM calls go straight from your machine to your provider — no Orkas server runs your agents.

LayerBuilt with
Desktop shellElectron — macOS / Windows from one codebase.
BackendTypeScript agent runtime in the app; can also drive local CLI coding agents (Codex, Claude Code, …) as subprocesses.
RendererVanilla HTML / CSS / JS via a sandboxed bridge — no framework lock-in.
StorageLocal JSON / JSONL files; in-memory hybrid (vector + keyword) search for the knowledge base.
ModelsBYO LLM keys via OAuth or API key; called directly, never proxied.
Open source: the full client is on GitHub — read the code, file issues, or contribute skills and agents.

# Data & privacy

Orkas is local-first: by default your agent configs, chats, knowledge bases, files, and memory live on your machine, and your encrypted keys and model traffic never go through Orkas. Lite's optional multi-device sync stores synced data on Orkas servers; the free edition sends limited usage analytics. Full details on the Security & Trust page.