The agent harness you can watch work.
A crew of AI agents on your own machine that research, write, code and run your routines — from the desktop or from Telegram, Discord and Slack. Every agent works in a living pixel station, so you can see exactly what each one is doing.
WATCH IT WORK
// every sprite is a real runDelegation
The overseer hands work to a specialist with team.dispatch. The worker runs on its own, with its own tools, and the result comes back to the lead.
Scripted replay of harness events — no model is running on this page.
- PLACED GEAR = REAL CAPABILITY
A signal dish grants web, a cabinet grants files, a workbench grants a terminal. Place one and every agent in that room draws on it.
- EVERY BAY = ANOTHER AGENT
Each one runs its own work in parallel. You direct from COMMS — they think, browse, build files, and deliver.
- THE FLOOR IS HONEST
An agent at a lit desk means a live run. The timer on screen is the real one.
// TRUTHFUL TELEMETRY
If it looks like an agent is working, it is.
HOW IT COMPARES
// same core, plus a station| STARNET | HERMES AGENT | OPENCLAW | |
|---|---|---|---|
| Runs on your own machine | YES | YES | YES |
| Chat from Telegram, Discord, Slack, Signal | YES | YES | YES |
| Persistent memory | YES | YES | YES |
| Browses the web, runs commands | YES | YES | YES |
| Watch every agent work, live | YES | — | — |
| Visual workflows between agents | YES | — | — |
| Per-job budgets and a cost ledger | YES | — | — |
Hermes Agent and OpenClaw columns show what each project’s own homepage describes (October 2026); — means it isn’t described there.
BOOT YOUR STATION
// three steps, about five minutes- 01Install StarNet
- 02Connect a model
- 03Send your first task
Choose the installer for your computer. Current stable version: v0.15.0. See what changed in the release notes. The station replay and docs follow current source, which may include features coming in the next release.
Bring your own model: OpenRouter, Anthropic, OpenAI, a ChatGPT sign-in, the Claude Code CLI, Gemini, Grok, Kimi, Mistral, DeepSeek, Groq — or Ollama, fully local.
git clone https://github.com/androoAGI/starnet.git
cd starnet && npm install && npm start
http://localhost:8787. Same app, same sidecar, no installer — keys stay in a local file instead of the OS keychain.Need help installing? Follow the quickstart or open troubleshooting. Public installers support Windows and macOS; Linux is not currently a supported public target.
FAQ
Is StarNet a harness or a visualizer for one?
A harness. StarNet’s own local process runs the agent loop, executes tools, routes work between agents, schedules runs, enforces budgets and keeps memory. The pixel station is how it shows you that work — it isn’t sitting on top of another agent framework.
Is StarNet free?
The software is MIT-licensed open source and free to run. Model usage costs whatever your provider charges — you bring your own API key (or OAuth account), and StarNet shows you real cost per run. No markup, no subscription required to use your own keys. If you'd rather not manage keys at all, StarNet Credits are an optional managed plan — never required.
Where does my data live?
On your machine. Station state, transcripts, memory, and ledgers live in your local StarNet workspace. Requests go to your model provider only when an agent runs, and to the wider network only when you explicitly grant a network tool or connector.
Can it spend money without me noticing?
No. Every conveyor line has a per-job budget ($2.00 unless you raise it, $50 ceiling) and an optional per-day cap; SETTINGS › SPENDING LIMITS adds per-run, per-agent and global caps the harness enforces against the real ledger, plus a $25-a-day limit (on by default) you can resume past or turn off. Autonomy is limited to a number of unattended jobs per day you choose. Use Stop in COMMS to interrupt the current task.
Why does my OS warn me during install?
SmartScreen can still warn while an Authenticode certificate builds reputation; inspect the publisher before proceeding. A public Mac DMG is required to pass Apple notarization, so a Gatekeeper trust failure should be reported rather than bypassed. Release checks do not prove that an installer works on your computer. The troubleshooting guide explains both cases.