← Back to research
·6 min read·opensource

Gastown

Gastown is Steve Yegge's open-source multi-agent orchestrator for running 20-30 Claude Code instances in parallel, built on his Beads data system.

Key takeaways

  • From Steve Yegge, former Google/Amazon engineer — orchestrates 20-30 Claude Code instances simultaneously
  • Built on Beads (225K lines of Go) — uses tmux as primary UI with seven specialized worker roles
  • Designed for expert users (Stage 7-8 on Yegge's AI coding evolution scale) — 'vibe coding' at scale

FAQ

What is Gastown?

Gastown is a multi-agent orchestrator that lets developers run 20-30 Claude Code instances in parallel using tmux and git-backed data storage.

Who created Gastown?

Steve Yegge, former senior engineer at Google and Amazon, created Gastown as his fourth attempt at an agent orchestrator in 2025.

What is the relationship between Gastown and Beads?

Gastown is built on top of Beads, Yegge's 225K-line Go library for git-backed data storage that serves as the control and data plane.

Is Gastown production ready?

No, Gastown is explicitly experimental. Yegge warns the codebase is under 3 weeks old and '100% vibe coded.'

Executive Summary

Gastown is Steve Yegge's experimental multi-agent orchestrator for running 20-30 Claude Code instances in parallel.[1] Built on his Beads data system, it uses tmux as its primary UI and implements seven specialized worker roles (Mayor, Polecats, Refinery, Witness, Deacon, Dogs, and the human Overseer). Gastown is explicitly for expert users comfortable with "vibe coding" at industrial scale.

AttributeValue
CreatorSteve Yegge
TypeOpen Source (Personal Project)
GitHub Stars9.3K
LanguageGo
HeadquartersN/A (Independent)

Product Overview

Gastown is an opinionated multi-agent orchestrator that Yegge describes as "Kubernetes for agents."[1] It's his fourth complete orchestrator of 2025, representing lessons learned from three failed predecessors. The system manages parallel Claude Code instances through a tmux-based interface with specialized worker roles.

The tool solves a specific problem: when running many AI coding agents simultaneously, work gets lost, agents conflict over git operations, and coordination becomes chaotic. Gastown provides structure through its "Town" and "Rig" abstractions.

Key Capabilities

CapabilityDescription
Multi-Agent Parallel ExecutionRun 20-30 Claude Code instances simultaneously
Seven Worker RolesMayor, Polecats, Refinery, Witness, Deacon, Dogs, Overseer
Merge QueueIntelligent conflict resolution via the Refinery role
Git-Backed StateAll state stored in git via Beads library
Tmux UITerminal multiplexer as primary interface

Worker Roles

RoleDescription
MayorMain agent, concierge and chief-of-staff
PolecatsEphemeral workers that swarm on tasks, produce Merge Requests
RefineryHandles Merge Queue, intelligently combines changes
WitnessWatches polecats, helps them get unstuck
DeaconDaemon beacon for system coordination
DogsTown-level monitoring and alerting
OverseerYou, the human operator

Technical Architecture

Gastown is built on two key components:[2]

  1. Beads — A 225K-line Go library providing git-backed data storage[3]
  2. Tmux — Terminal multiplexer serving as the primary UI

Core Concepts

ConceptDescription
TownYour HQ directory (e.g., ~/gt) managing all rigs
RigA single project/repo under Gastown management
MR (Merge Request)Work unit produced by polecats
MQ (Merge Queue)Ordered queue of MRs awaiting integration

Key Technical Details

AspectDetail
DeploymentLocal (CLI + tmux)
Model(s)Claude Code (and compatible CLIs)
DependenciesBeads, tmux, Go runtime
Open SourceYes (GitHub)

Degradation model: Gastown degrades gracefully — every worker can operate independently, and you can run partial configurations or even "no-tmux" mode.


Strengths

  • Scale — Enables 20-30 parallel agent instances, far beyond manual management
  • Intelligent merging — Refinery role handles the complex "monkey knife fight" of parallel git operations
  • Graceful degradation — Works in partial configurations, doesn't require full orchestration
  • Expert pedigree — Built by Steve Yegge with decades of Google/Amazon engineering experience
  • Open exploration — Tackles the MAKER benchmark problem (20-disc Hanoi towers) that defeats other approaches[4]
  • Built on proven foundation — Beads already has tens of thousands of users

Cautions

  • Explicitly experimental — Codebase under 3 weeks old at launch, "100% vibe coded"
  • Expert-only — Requires Stage 7-8 on Yegge's AI coding evolution scale
  • High cost — Multiple Claude Code accounts needed; "cash guzzler"
  • Tmux required — Must learn tmux to use effectively
  • Beads lock-in — No alternative backend; must adopt Yegge's full stack
  • Chaotic by design — "Work can be chaotic and sloppy" — bugs get fixed multiple times, designs go missing

Pricing & Licensing

TierPriceIncludes
GastownFreeOpen source
Claude Code$20-100/moRequired for each account

True cost: Yegge mentions needing multiple Claude Code accounts and expecting to need a third within a week of launch. Heavy users should budget $100-300+/month.

Licensing model: Open source (personal project)


Competitive Positioning

Direct Competitors

CompetitorDifferentiation
RalphRalph is simpler (bash loop); Gastown is full orchestration system
TemboTembo is commercial with enterprise features; Gastown is experimental/personal
Manual tmuxGastown adds structure, roles, and merge queue to raw tmux management

When to Choose Gastown Over Alternatives

  • Choose Gastown when: You're already at Stage 7-8 of AI coding evolution, comfortable with chaos, and want to push the boundaries of parallel agent execution
  • Choose Ralph when: You want the simplest possible approach (bash loop) without the orchestration complexity
  • Choose Tembo when: You need enterprise features, team support, and stability

Ideal Customer Profile

Best fit:

  • Expert developers already managing 3-5+ parallel Claude Code instances
  • Engineers comfortable with tmux and terminal-first workflows
  • Researchers exploring multi-agent coordination
  • Developers willing to accept chaos and instability for throughput

Poor fit:

  • Anyone not already at Stage 6+ of AI coding adoption
  • Teams needing stability and reliability
  • Developers uncomfortable with vibe coding philosophy
  • Organizations requiring enterprise support or compliance

Viability Assessment

FactorAssessment
Financial HealthN/A — Personal open source project
Market PositionPioneer — Pushing boundaries of multi-agent orchestration
Innovation PaceRapid — Fourth orchestrator in one year
Community/EcosystemGrowing — 9.3K stars, builds on Beads (tens of thousands of users)
Long-term OutlookUncertain — Depends on Yegge's continued interest

Gastown represents cutting-edge experimentation in multi-agent orchestration. It's not production-ready and explicitly warns users away, but for the right expert users, it demonstrates what's possible.


Bottom Line

Gastown is the most ambitious open-source multi-agent orchestrator available, enabling 20-30 parallel Claude Code instances with sophisticated role-based coordination. It's explicitly experimental and only appropriate for expert users already pushing the limits of AI-assisted coding.

Recommended for: Expert developers (Stage 7-8) who want to explore the frontier of multi-agent orchestration and are comfortable with instability.

Not recommended for: Anyone seeking production stability, enterprise features, or a gentle learning curve.

Outlook: Gastown represents where multi-agent orchestration is heading. The concepts (specialized roles, merge queues, graceful degradation) will likely influence commercial tools, even if Gastown itself remains experimental.


Research by Ry Walker Research • methodology