A multi-agent orchestration system for Claude Code, GitHub Copilot, and other AI coding agents with persistent git-backed work tracking, automatic task decomposition, and Bors-style merge queues.
Gas Town is an orchestration system designed for multi-AI coding agent collaboration, maintained by the gastownhall organization. It is primarily written in Go (94.8%) and exposes a CLI tool named gt. The latest release is v1.0.0 under the MIT License.
Core Mechanisms#
The system uses git as its persistence layer — all agent state is written via git hooks into the repository, solving the fundamental pain point of AI agents losing context after restarts. The orchestration model has three layers: The Mayor (AI coordinator, defaulting to a Claude Code instance) receives user goals and automatically decomposes them into tasks; Rigs encapsulate individual git repos and their associated agents; Polecats serve as worker agents with persistent identities but temporary sessions.
Workflow and Merging#
Tasks are packaged and assigned via Convoys (work tracking units) with autonomous stall detection and smart skip capabilities. Beads provides git-backed issue tracking (format: gt-abc12). Molecules define repeatable workflow templates in TOML (Formulas), supporting root-only wisps and poured wisps modes. Refinery implements a Bors-style bisecting merge queue for batch verification and PR merging, with automatic isolation of failed MRs. Escalation provides P0/P1/P2 severity-graded issue routing.
Monitoring and Observability#
A three-tier monitoring system: Witness (per-rig lifecycle management) → Deacon (cross-rig patrol) → Dogs (infrastructure maintenance). Provides an htmx-driven Web Dashboard (single-page, auto-refresh) and a three-panel terminal TUI (Activity Feed with Agent Tree / Convoy Panel / Event Stream / Problems View). Full-chain OpenTelemetry telemetry outputs structured logs and metrics to OTLP-compatible backends.
Cross-Instance Collaboration#
Wasteland Federation connects multiple Gas Town instances via DoltHub, supporting cross-instance task publishing, claiming, and a portable reputation mechanism (specific DoltHub instance URL to be confirmed).
Agent Runtimes and Scheduling#
Ten built-in presets: claude, gemini, codex, cursor, auggie, amp, opencode, copilot, pi, omp. Custom agent runtimes can be added via gt config agent set. Scheduler acts as a capacity controller to prevent API rate limit exhaustion, with configurable concurrency limits.
Deployment#
Supports Homebrew (brew install gastown), npm (@gastown/gt), Go source compilation, manual builds, and Docker Compose deployment across macOS, Linux, and Windows. Prerequisites include Go 1.25+, Git 2.25+, Dolt 1.82.4+, beads (bd) 0.55.4+, sqlite3, and tmux 3.0+ (recommended). Also includes Nix support (flake.nix), shell completions (Bash/Zsh/Fish), and a plugin system (plugins/ directory).
Unconfirmed Information#
- Steve Yegge's specific role (founder/contributor/former maintainer) is not explicitly stated in the README
- The Go 1.25+ version requirement may be forward-looking or a typo (Go 1.25 was not yet officially released at time of collection)
- No publicly available user case studies or production deployment reports
- No independent website or specific DoltHub Wasteland instance URL provided