Gas Town#
Open-source multi-agent orchestration system for AI coding agents. Written in Go, distributed as a single CLI binary (gt). Coordinates 20-30+ concurrent agents with persistent work tracking via git worktrees.
Key Facts#
- Repository: https://github.com/gastownhall/gastown
- Language: Go 1.25+
- License: MIT
- Install:
brew install gastown/npm install -g @gastown/gt/go install - Dependencies: Git 2.25+, Dolt 1.82.4+, beads (bd) 0.55.4+, tmux 3.0+ (recommended)
- Supported runtimes: Claude Code, Codex, GitHub Copilot, Gemini, Cursor, Auggie, Amp, OpenCode, Pi, OMP
Architecture#
Hierarchical workspace model:
- Town → workspace root
- Rig → project container (git repo + agents)
- Polecat → worker agent (persistent identity, ephemeral sessions)
- Hook → git worktree for agent state persistence
- Convoy → work tracking unit (bundles beads/issues)
- Refinery → Bors-style merge queue per rig
- Wasteland → federated work network via DoltHub
Monitoring#
Three-tier watchdog: Daemon → Boot → Deacon → Witnesses/Refineries. Detects stuck agents (GUPP violations), zombies, and stalls. Escalation routing: Deacon → Mayor → Overseer.
Differentiators#
- Git worktrees as the persistence primitive (not databases, not memory stores)
- Agent-agnostic — any CLI-based AI tool can be a runtime
- Built-in merge queue with bisecting failure isolation
- Federated cross-workspace coordination (Wasteland)
- Session discovery (Seance) for context continuity across agent restarts
- OpenTelemetry observability built-in