I built RustyClaw, a system for running multiple AI coding agents as persistent daemons that collaborate in teams.
The idea started from OpenClaw's approach to multi-agent coordination, but I wanted something faster, more reliable, and fully open source — so I rebuilt it from scratch in Rust.
How it works: you define agents with different roles (@coder, @review, etc.), group them into teams, and the system handles routing and handoffs automatically. When @coder finishes implementing something, it mentions @review in its response, and the queue processor detects it and routes the message. No central orchestrator LLM — agents coordinate through a simple file-based queue.
Key design decisions:
Provider-agnostic — works with Claude Code, OpenAI Codex CLI, and OpenCode. Mix and match models per agent
tmux daemon — everything runs in a persistent tmux session. Attach, detach, check logs, no web UI required (though there's an optional WASM visualizer)
Identity files — each agent has SOUL.md, IDENTITY.md, MEMORY.md in its working directory. The agent's personality and role are defined in plain markdown, not prompt engineering hidden in code
Supports Discord and Telegram as input channels, or just use the CLI directly.
jurgsieg•1h ago
The idea started from OpenClaw's approach to multi-agent coordination, but I wanted something faster, more reliable, and fully open source — so I rebuilt it from scratch in Rust.
How it works: you define agents with different roles (@coder, @review, etc.), group them into teams, and the system handles routing and handoffs automatically. When @coder finishes implementing something, it mentions @review in its response, and the queue processor detects it and routes the message. No central orchestrator LLM — agents coordinate through a simple file-based queue.
Key design decisions:
Provider-agnostic — works with Claude Code, OpenAI Codex CLI, and OpenCode. Mix and match models per agent
tmux daemon — everything runs in a persistent tmux session. Attach, detach, check logs, no web UI required (though there's an optional WASM visualizer)
Identity files — each agent has SOUL.md, IDENTITY.md, MEMORY.md in its working directory. The agent's personality and role are defined in plain markdown, not prompt engineering hidden in code Supports Discord and Telegram as input channels, or just use the CLI directly.