Orkas-AI/Orkas

Open-source multi-agent AI desktop client — build and command your AI agent team through conversation. A commander LLM dispatches sub-agents in parallel or in series; agents self-evolve via reflection and skill crystallization. Local-first, BYO LLM keys (Claude · OpenAI · Gemini · DeepSeek · Kimi · GLM · Qwen). macOS / Windows / Linux.

What it solves

Orkas provides a local-first way to manage a workforce of specialized AI agents without needing to write orchestration code or rely on cloud-hosted SaaS platforms. It solves the problem of coordinating multiple AI agents for complex tasks while ensuring that sensitive data, API keys, and conversation histories remain on the user's local machine.

How it works

Orkas operates as a desktop application featuring a central "Commander" agent that acts as the orchestrator. The Commander breaks down high-level goals into executable paths and dispatches specific tasks to specialist agents.

Key technical mechanisms include:

  • Visibility Slicing: Agents only see the parts of the conversation relevant to them, reducing token usage and protecting private context.
  • Structured Dispatch: The Commander uses specific tool calls to assign work, ensuring tasks are queued and executed in a controlled, turn-based sequence.
  • Self-Evolution: Each agent maintains its own memory and competence files (COMPETENCE.md and LEARNING_STRATEGIES.md), reflecting on completed tasks to improve its performance over time.
  • Local-First Architecture: All files, keys, and knowledge bases are stored locally; the app connects directly to LLM providers (or local endpoints) without proxying data through external servers.

Who it’s for

Orkas is designed for users who want a multi-agent AI workforce with a graphical user interface rather than a coding framework. It is ideal for those who need to automate recurring reports, translate product specs into development tasks, or perform local data analysis while maintaining strict data privacy.

Highlights

  • Local-First Privacy: Conversations, files, and API keys never leave the local disk.
  • Commander-Led Orchestration: A single interface to direct a team of agents without writing flowcharts or code.
  • Extensible Ecosystem: Ability to integrate external CLI coding agents like Claude Code, Cline, and OpenCode.
  • BYO LLM: Supports a wide range of providers including Claude, OpenAI, Gemini, DeepSeek, and local model endpoints.
  • Self-Improving Agents: Agents can crystallize successful problem-solving methods into private, reusable skills.