A pixel-art station where the floor plan is load-bearing

What it does StarNet is a desktop harness for running a crew of AI agents that make real model calls, run real tools, and incur real cost — it explicitly refuses to animate a simulation. You create agents with distinct classes, personas, and loadouts; the harness runs several concurrently, each with its own workspace, transcript, memory, and bounded permissions. Under the hood it’s a Tauri shell, a vanilla-JS frontend, and a local Node sidecar built on core modules only, streaming events over localhost NDJSON/SSE. Bring an OpenRouter key or provider OAuth, or run keyless against a local Ollama model.

The interesting bit The station is the permission model, not a skin over it: a room is a capability-scoped team, a hallway is an authorized handoff lane, a placed object is a real capability grant. The project’s stated law — the interface must never assert state the harness cannot prove — reads as a pointed rebuttal to agent dashboards that render fiction. Spend, budgets, and run history persist on disk and are shown as-is.

Key highlights

• Local-first: station state, transcripts, memory, and ledgers stay on your machine; only provider calls leave it.
• BYOK done properly — keys live in the OS keychain via the sidecar, never in the frontend.
• Wire agents to Telegram, Discord, Slack, Signal, or Matrix; Night Shift lets them work unattended inside an explicit, logged leash.
• Task Briefs turn ambiguity into one concrete question with options instead of a silent wrong guess; finished work lands in the OUTBOX as real files, not chat scrollback.
• MCP connectors, cron schedules, reusable skills and recipes, plus an importer that mints agents from existing OpenClaw or Hermes homes.

Caveats

• Windows is the most-tested target; macOS has less real-world coverage, and Linux builds are internal artifacts, not a supported public release.
• MIT covers the code only — the StarNet name, logo, and artwork stay with the author, so forks must fully rebrand.
• The keyless Ollama route is free but, by the README’s own admission, slower and rougher on long tasks.

Verdict Worth a look if you want a self-hosted, BYOK multi-agent harness where permission boundaries are walls you can see instead of config you squint at. Skip if you need Linux support or consider pixel art a tax.