Welcome to CHAOS-AI v4.3.0
Introducing CHAOS — the Context Hydrated Agentic Orchestration System. 36 agents, 134 MCP tools, and a shared context layer that cuts token usage by 50%+.
What is CHAOS-AI?
CHAOS (Context Hydrated Agentic Orchestration System) is a local-first, provider-agnostic multi-agent AI workspace for the terminal. While single-agent tools reset every session and re-read the same files on every run, CHAOS coordinates 36 specialized agents through a shared persistent context layer — so every agent starts informed, and what one learns, all benefit from.
The result: faster development, 50%+ fewer tokens spent, and AI-assisted work that actually understands your project.
Why Multi-Agent?
Single-agent AI assistants hit a wall when projects grow complex. Context windows fill up. The same architectural questions get answered from scratch every session. Agents repeat each other's work without knowing it.
CHAOS solves the coordination problem. A PM engine dispatches agents to real subprocesses via a PostgreSQL task queue. A shared event bus keeps them synchronized. A three-database context layer — SQLite, Praxis Store, and knowledge graph — gives every agent access to project history, coding conventions, and recent decisions.
What Ships in v4.3.0
- 36 specialized agents — Python, TypeScript, Go, Rust, Java, C#, Swift, C, C++, R, plus quality, infra, data science, mobile, docs, and management agents
- 134 MCP tools — expose the full CHAOS capability surface to any MCP-compatible editor (Claude Desktop, VS Code Copilot, Cursor, Windsurf)
- 80 skills + 72 slash commands — the full development lifecycle, fuzzy-matched from plain language
- 6 provider adapters — Claude Code, GitHub Copilot CLI, Gemini CLI, Aider, OpenRouter, Ollama
- Cross-clone git sync — live event bus keeps parallel agent instances aligned across worktrees
- Web dashboard — real-time agent monitor, token analytics, event stream, and security overview
Getting Started
CHAOS is installed from source and runs via the cw CLI:
git clone https://github.com/luvisaisa/claude-workflow
cd claude-workflow
pip install -e .
cw --version
Start the MCP server to connect your editor:
cw mcp
# or for HTTP transport:
cw mcp --transport streamable-http
Run the PM agent to start orchestrating:
cw run pm
Check out the documentation for the full setup guide, or see the benchmarks for the numbers.
What's Next
We are building in the open. Follow along on GitHub and join the conversation.