StarNet is a local-first desktop agent harness that runs a crew of real AI agents and presents them as a pixel-art space station. It is aimed at developers and tinkerers who already call language models with their own API keys and want more structure than a chat window gives them: several agents working at once, each with its own workspace and its own bounded permissions. The station is not a skin over a log. The repository is explicit that the picture is a projection of live runtime state, and that the layout you draw is the workflow the agents actually run.
What it does
You create agents with distinct classes, personas and loadouts, and then arrange them in the station. The arrangement carries meaning, and the project states the mapping as a product contract: a room is a capability-scoped team, a hallway is an authorized handoff lane, and a placed object is a real capability grant. Dropping an object into a room is how an agent gets a permission, not a decoration step you do afterwards.
Several agents can run concurrently. Each run is described as genuinely distinct and bounded, with its own workspace, transcript, memory and permission set, so two agents working in parallel are not sharing one context. You scale up by placing more bays or summoning specialists rather than by opening more tabs.
Around that core sit a few practical pieces:
- Bring your own models: you paste your own key, and the project's topics name OpenRouter among the providers it targets.
- Chat bridges, so you can message your crew from Telegram, Discord or Slack instead of sitting in front of the app.
- An OUTBOX, where finished work lands as real files on disk.
The README is equally clear about the flip side: the harness performs real model calls, real tools and real cost. Nothing is animated as a simulation, which also means nothing is free.
How it works
StarNet is a Tauri desktop application, mostly JavaScript, distributed for Windows and macOS. The station view is rendered from live harness state, so what you see reflects which agents exist, what they are permitted to touch and what they are doing. Because the layout is the permission model, editing the map is the configuration interface, and there is no separate policy file to keep in sync with the drawing.
Local-first here means the program runs on your machine with your key, and the project ships a privacy document alongside the readme to describe what that entails. The corollary is that model spend goes on your own provider account.
Getting started
Prebuilt releases live in a companion repository, linked from the readme as a download for the latest release, with a separate install guide for platform specifics. If you would rather build it yourself, the readme has a run-from-source section, and there is a docs index and a contributing guide for people who want to go further than clicking around. The license is MIT.
The natural first session is small: paste a key, make one agent, give it a room and one object, and see what it can and cannot reach. The station metaphor only pays off once you have a second agent to hand work to.
When to use it / when not
Reach for it when you want multiple bounded agent runs with a visible permission model, when you want those runs on your own hardware with your own key, and when you would rather see agent capability as a floor plan than infer it from a configuration file. The chat bridges and the OUTBOX make it plausible as something you leave running and check on, rather than a demo you drive continuously.
Skip it if the platform badge rules you out — the readme advertises Windows and macOS only. Skip it if you want something to import into your own program; this is a standalone desktop app, not a library. Skip it if you expect the agents to be cheap, since the project's own framing is real calls and real cost. And treat it as young: the repository was created in June 2026 and stands at a few hundred stars, so expect the interfaces and the contract wording to keep moving.
Anyone building or supervising multi-agent work should take StarNet seriously for one reason above the artwork: it turns permissions into something you can see and place. Most harnesses let capability grants hide in configuration, where a mistake is invisible until an agent does something you did not intend. Making a room a scope and an object a grant is a real design choice, and a game-like surface is an unusually honest way to expose it. The pixel art will decide whether people try it; the permission model is what would make them stay.