Gas Town

Gas Town is the orchestrator for the person who wants to hand a coordinator agent a pile of tickets and come back to merged code. It drives Claude Code by default, with presets for GitHub Copilot, Codex, Gemini, Cursor, Kiro, Amp, OpenCode and Pi, and the trait that decides it is unattended operation: workers run in tmux while you are elsewhere, and the merges happen without you.

You need enough independent work that a human dispatcher is the bottleneck. Pick it if you do and enjoy Steve Yegge's vocabulary. Pick Symphony for a Codex-plus-Linear loop, and Claude Squad if you want to watch every pane.

The failure mode begins at install. The README requires Git 2.20 or newer, Go 1.26.2 or newer, Beads 0.57.0 or newer, sqlite3, the ICU4C development headers and tmux 3.0 or newer, and the documented setup line runs gt doctor --fix before anything works, which is a tool admitting the install will need repair. After that, workers run on the host with no container layer, many at once.

Use the Docker Compose route, which needs only Docker on the host. What it does right: a per-rig Witness detects stuck workers and a background Deacon patrols every rig, so a hung agent is noticed by the system rather than by you.

The mechanisms are named and documented. 1. Identity: a Polecat worker keeps its identity and history while its session is discarded after each task, so context does not accumulate across jobs. 2. State: work items are Beads, stored in git, bundled into Convoys, so the tracker has the same history as the code. 3. Integration: when a worker finishes with gt done, the Refinery batches merge requests, runs verification gates and merges with Bors-style bisection, isolating the commit that breaks the build.

No benchmark is published. The observation: this is the only orchestrator here whose merge step is a queue rather than a hope.

Gas Town Hall is a GitHub organisation, not a corporation, and the asset is Steve Yegge's name: a known essayist whose posts move engineers, attached to 17,922 stars since a December 2025 release. There is no pricing, no hosted tier and no funding, and the tool consumes other vendors' subscriptions, so the revenue accrues to Anthropic, not to the project.

Moat: the author's audience, which is also the risk when the author's attention moves. Likely outcome: not an acquisition but an inspiration, the Beads and merge-queue ideas reappearing in funded products. Position: read it, borrow from it, do not depend.

The demo is a town of agents merging overnight. Procurement: there is no vendor, no seat and no contract, so the cost is the meter, and the design multiplies it, one coordinator plus several workers per engineer against the subscription each already holds. There is no SSO, no central log, and the setup command gt install ~/gt --shell --git modifies the engineer's shell and git configuration, which our endpoint team will ask about.

It does not run in CI. Onboarding is an afternoon for a tmux-literate engineer and longer for the rest. Not yet; one shared deployment with a cost cap, then a pilot of five.

MIT and Go, installed with go install for both gt and the bd issue tool, so I can build it from a checkout. The part I like is the propulsion principle: git hooks drive the state machine, so coordination is in a repository I can read, diff and roll back rather than in a database I cannot. The preset list goes past the famous names to Auggie and OMP, and adding my own is a config entry.

No MCP and no model setting, since the agents bring both. Forking is go build with a new name. Eccentric, documented, and entirely inspectable.