What they did
Sanyal keeps a growing set of specialised Claude Code sub-agents, among them a market sizing agent and trading agents that study stocks, spread across project folders. He lost track of them in the desktop app, so he built Agent City, an isometric pixel town that runs on his Claude Pro plan. The app scans his project folders for sub-agent markdown files, draws each folder as a fenced lot and each agent as a character. Clicking one and typing a task starts a real Claude Code session through the CLI with streaming JSON output, and a meter shows tokens in and out. A Mayor agent, a Claude Code session with a small MCP server that can list agents and ask any of them a question, coordinates multi-agent jobs. Claude Code wrote most of the Node, Express and React code from his written design and a step-by-step plan, with tests for each step.
How it went
He says it now replaces the Claude desktop app for him at work and at home, and released it as open source under MIT. New agents can be hired from the town, with Claude drafting the instructions from a short description. He also added a chat that can switch to local Llama and Qwen models when session limits run out. He gives no numbers on build time, cost or how much the orchestrator helps.
Worth knowing
It is Mac only and needs Node 20 and a signed-in Claude Code install. Every agent job is a full Claude Code session, so jobs draw on the same plan limits as using the CLI directly; the token meter exists for that reason. The server binds only to localhost, but prompts still go to Anthropic.