Gastown — Multi-Agent Workspace Manager

Introduction

Gastown is an open-source workspace manager that orchestrates multiple AI coding agents within a single project. It provides shared context, task routing, and session management so that concurrent agents do not duplicate work or produce conflicting edits.

What Gastown Does

• Manages multiple AI agent instances working on the same codebase concurrently
• Routes tasks to agents based on specialization and file ownership
• Provides shared context so agents see each other's in-progress changes
• Maintains session history across agent restarts for long-running workflows
• Offers a dashboard for monitoring agent activity and resolving merge conflicts

Architecture Overview

Gastown runs an event-driven workspace server that maintains a shared state graph of the project. Each agent connects as a client and registers capabilities. The server handles file locking, change propagation, and context windowing. A task planner decomposes objectives into sub-tasks and assigns them to agents. Inter-agent communication uses structured message channels, keeping the system loosely coupled and provider-agnostic.

Self-Hosting & Configuration

• Requires Node.js 18+ and Redis for messaging and state coordination
• LLM provider credentials are set in .env; supports OpenAI, Anthropic, and Ollama
• Workspace scope and file ownership rules live in gastown.config.yaml
• Dashboard port is configurable and can sit behind nginx for team access
• Concurrency limits and context budgets are adjustable per workspace

Key Features

• Provider-agnostic orchestration that works with any LLM backend
• File-level locking to prevent conflicting edits across agents
• Task decomposition that breaks objectives into parallelizable sub-tasks
• Session persistence across restarts and network interruptions
• Real-time dashboard with agent status and unified diff view

Comparison with Similar Tools

• CrewAI — general multi-agent framework; Gastown is purpose-built for code workspaces with file locking
• AutoGen — multi-agent conversations; Gastown adds workspace-aware context sharing and monitoring
• OpenHands — agent runtime environment; Gastown focuses on coordination and conflict resolution
• Cursor — single-agent IDE; Gastown orchestrates multiple independent agents outside the IDE

FAQ

Q: How many agents can run concurrently? A: Defaults to 8 per workspace, adjustable in config. The practical limit depends on system resources and API rate limits.

Q: Does it require a specific LLM provider? A: No. It is provider-agnostic with built-in adapters for OpenAI, Anthropic, and Ollama.

Q: How are conflicting edits handled? A: File-level locking and change propagation notify agents of concurrent work. Unresolved conflicts surface in the dashboard for human review.

Q: Can I use it with my existing IDE? A: Yes. Gastown runs as a background service. Changes appear in your IDE's git diff view as agents make them.

Sources

• https://github.com/gastownhall/gastown