androoAGI/starnet

The problem

Running several LLM agents on real tasks usually means juggling scripts, terminals, and API dashboards, with no clear picture of what is running, what it costs, or what it is allowed to touch. StarNet turns that into a single desktop station: you place agents in rooms, and the layout you draw determines which capabilities and handoff lanes they get.

What's different

The crowded space here is generic agent frameworks and orchestration libraries, where the topology of a multi-agent system lives in config files and logs. StarNet's bet is that the spatial arrangement itself is the permission model: a room is a capability-scoped team, a hallway is an authorized handoff, a placed object is a real grant. It is also unusual in making each concurrent agent a distinct bounded run with its own workspace and transcript, and in displaying spend and run history from disk rather than a synthesized animation. The BYOK model (OpenRouter key or Anthropic/OpenAI/Google sign-in, keys in the OS keychain) and the overlay of messaging channels onto the same station are useful, though not unique on their own.

Who it's for

Best for individual developers or small teams who already run agents against real APIs and want several working in parallel with visible budgets and approvals, on Windows or macOS desktops. It also suits people who prefer a game-like interface over a web dashboard. Look elsewhere if you need headless Linux deployment, a CI-oriented orchestration layer, or strict reproducibility; the public release train is Windows/macOS only and Linux builds are internal.

Cost of adoption

Desktop builds are signed and notarized per the README, but the project labels itself an early release with Windows the most-tested target and less macOS coverage. Source runs need Node.js 18+ (22 matches CI), with Rust and Tauri prerequisites only for the desktop shell. Local-model use via Ollama is free but the README itself warns small models handle long tasks poorly and an 8B model needs roughly 10 GB of graphics memory. The MIT license covers code only; the name, logo, and artwork are reserved. Expect to re-enter API keys when importing from OpenClaw or Hermes.

Verdict

StarNet is a young project, a little over four months old, and its 1,598 stars with a +1,491 monthly board appearance suggest one sharp spike rather than a long climb; no Hacker News discussion and no daily trending presence mean outside validation is thin. The concept — spatial permissions as a literal contract — is genuinely distinctive and the README is unusually candid about what is and is not real. But the value hinges on the harness holding up under concurrent real workloads, which is exactly what is hardest to prove early. Treat it as an interesting prototype to try on a secondary machine, not as the backbone of a production agent workflow yet.