Pause, whisper, and reset a town of agents

What it does GOD is a local-first operator console for running and steering societies of LLM agents. It bundles a React frontend, FastAPI backend, and out-of-process agent runtime into a single stack that runs entirely on your machine, talking to any OpenAI-compatible endpoint. You pick or build a scenario, generate residents via a browser wizard, then step into a live control room to watch and intervene.

The interesting bit The project treats simulation like a replayable game engine rather than a batch script. Every step is stored locally in SQLite, so you can scrub back and forth through time, and injected instructions hit the agents on their very next turn—no restart required.

Key highlights

• Real-time intervention: pause, fast-forward, or inject commands into the next agent step mid-run.
• Targeted messaging: ask one resident, a group, or the whole town without leaving the map.
• No-code setup wizard: configure models, draft scenarios, and generate agent profiles from the browser.
• Local-first architecture: experiment files, replay store, and runtime state stay on your machine; only the LLM endpoint is external.
• Built-in god_town scenario: a 10-resident slice-of-life world with 65 location-scoped interactions.

Caveats

• Still alpha; expect rough edges and moving parts.
• Only one experiment can be active at a time, tracked in .god/current_experiment.json.
• Requires an external OpenAI-compatible API; there is no built-in model.

Verdict Worth a look if you are building or studying multi-agent social simulations and need human-in-the-loop steering. Skip it if you want fully autonomous, headless agent orchestration without a map or replay UI.