The starnet ai agent is a free, open-source desktop agent harness that turns your AI agents into little pixel-art characters living on a space station, where every room, hallway and object maps to something real the agent can actually do. I set one up on camera for the Julian Goldie Podcast, and this is the plain walkthrough: what it is, how the first agent comes to life, what each part of the station does, and whether it's worth your time.
📺 Watch: StarNet: FREE Pixel AI Agent Office!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
Short version: it's fun, it's the smoothest office-style agent interface I've tried, and it's early. Keep reading if you want the detail.
What Is StarNet?
Per the README on the official repo, StarNet is "a living pixel-art station where real AI agents do real work." It's built by Andrew Sims (androoAGI). It runs on your own machine, it's local-first, and you bring your own key or model.
The key idea is that the station isn't decoration. Per the README, the station is a projection of live runtime state:
- A room is a capability-scoped team.
- A hallway is an authorised handoff lane between agents.
- A placed object is a real capability grant.
You start with one agent. Then you place bays or summon specialists. Each one is a distinct, bounded agent run with its own workspace and permissions. So when you look at the station, you're looking at what your agents can do and who can hand work to whom.
The README also says StarNet "does not simulate revenue, completed work, model activity, or spend". That matters. A lot of pixel-office projects are eye candy. This one says the picture reflects real state.
If you want to go deeper on AI agents like this and the setups I actually use day to day, the AI Profit Boardroom has full Hermes and AI agent courses plus my Agent OS, for $69/mo.
How I Set Up My First StarNet AI Agent
If you still need to install it, my StarNet GitHub install guide covers the official repo, the installers and running it from source. Once it's open, here's what the first-run flow looked like for me.
Name it and give it a personality
First you name your agent. Then you pick an appearance and a personality. The options were warm, dry, composed or unhinged. Pick whatever you like talking to, because you'll be chatting with it a lot.
Choose how much control it gets
Next it asks how much power the agent should have. I chose "ask for approval" over "full power". That means the agent checks with you before it does anything that matters. For a first agent, that's the sensible choice. You can loosen it later once you trust it.
Connect a brain
Then you connect a model. You can sign in with an existing subscription (Claude Code was one of the options I mentioned) or use a provider like Kimi or Grok. I signed in with Grok in one message and picked Grok 4.6, which I prefer to 4.7. You also pick a reasoning level: low, medium, high or max.
Per the README, you can also paste an OpenRouter key, sign in with Anthropic, OpenAI or Google accounts, or run it free on a local model with Ollama. Keys live in your operating system's keychain, never in the frontend.
It wakes up
Once the brain is connected, the agent "wakes up" with a little animation. That's your first crew member on the station.
📺 Watch: Hermes3D Is INSANE: AI Agents Now Have a 3D Office
A Tour of the Station
This is where StarNet stands out. Everything is visual, but each panel does a job.
Station view
You can zoom in and out, switch frames, and use four docks. The agent levels up the more you use it. It's like a Tamagotchi for an AI agent. That sounds silly, but it does make you want to come back and use it.
Agent settings panel
Click your agent and you get its settings: model, personality, instructions, purpose, access (web and local files) and approval prompts. There are tabs for brief, growth, history and memories. So you can see what it knows, what it's done and how it's grown.
Chat
Chat is where you give it work. I told mine I wanted SEO help for aiprofitboardroom.com. It picked that up and ran with it. I did notice a slight delay when sending messages, so don't expect instant replies every time.
Per the README, StarNet also does task briefs, which turn a vague ask into one concrete question with options. And deliverables land in an OUTBOX folder as real files, not just chat text.
Widgets
The widgets area shows connected apps, pinned widgets and a daily brief. It's a quick dashboard for what your agent is plugged into and what's happening today.
App catalog
The app catalog is big. You've got Gmail, Notion and GitHub. Most sign in directly, and some need an API key. There are also real-world categories, like finding flights and sending physical letters, plus web search. Per the README, it supports MCP connectors and messaging through Telegram, Discord, Slack, Signal and Matrix.
Skill market and skill library
The skill market has things like ad copy testing and calendar scheduling. Some are built in, some you install, some come from the community. Then there's the skill library, which is a set of procedures you toggle on and off: a decision framework for hard problems, planning, humanising content, and handling hard conversations.
Crew and recruiting
Under Crew you can recruit specialists: a strategist, a chief of staff, a researcher and a marketer. Here's the one thing to get right:
- Deploy overwrites your current agent's setup with the new role.
- Summon adds a second agent beside it.
If you want a team, use summon. If you click deploy by mistake, you replace the agent you already set up.
Tasks, quests, systems and builds
There are also menus for tasks, quests, systems and builds. Per the README, there's a Night Shift feature too, where agents keep working on cron schedules with guardrails, plus push-to-talk voice. Spend, budgets and run history are saved on disk, so you can see what it's costing you.
📺 Watch: Hermes AI Agents Just Went Portable
Honest Verdict on the StarNet AI Agent
Here's my straight take. I don't think there's a massive amount of practical use cases for this yet, but it's fun.
It's the smoothest office-style agent UI I've tried. It feels like a much more fun version of Paperclip. If you've seen my Paperclip and Hermes agent breakdown, you'll get the comparison right away. It also reminded me of the first time I tried Hermes: that feeling of "oh, this is where things are heading."
If you like the visual-office idea, it sits in the same family as the Hermes pixel office. The difference is that StarNet ties the art to real permissions and real runs.
What to keep in mind:
- It's an early release. Per the README, Windows is the most-tested platform and macOS has less real-world coverage.
- Replies can lag a little. I saw a slight delay on sends.
- Practical value is still growing. For serious daily work, I'd still lean on my main agent setup.
If you want help working out which agent setup actually fits your business, book a free AI strategy session and we'll map it out.
Does StarNet Work With Hermes?
This was the most common question in the comments. Yes. Per the README, StarNet can import existing agents from OpenClaw or Hermes installs, including persona, instructions, memory and model. You just re-enter your API keys after import.
So if you've already built a Hermes agent you like, you don't start from scratch. You bring it onto the station. Viewers also asked for more tutorials on how it all works together, so expect more on that.
You can also plug open-source projects like StarNet into my Agent OS with the zip setup. The Agent OS guide explains how that works, and AIPB members get it.
Who Should Try StarNet?
Try it if:
- You're curious about AI agents and want a friendly, visual way in.
- You already run Hermes or OpenClaw and want to see your agents in a new interface.
- You like experimenting with multi-agent teams and want to see handoffs and permissions laid out visually.
- You want a free option. Run it with Ollama and a local model and you pay nothing for the model.
Skip it for now if:
- You need a battle-tested tool for client work today.
- You're on Linux and want a prebuilt installer (there isn't a supported public Linux release).
- You just want results and don't care about the interface.
If you're comparing models to run inside tools like this, check Goldie Bench for how I test them.
And if you want the full courses on Hermes, AI agents and my Agent OS in one place, join the AI Profit Boardroom for $69/mo.
FAQ
Is StarNet free?
Yes, StarNet itself is free. The code is MIT licensed and you bring your own key. Model costs depend on what you connect. Per the README, you can run it free with a local model by installing Ollama and picking Ollama as the provider.
Does StarNet work with Hermes?
Yes. Per the README, StarNet imports existing agents from Hermes or OpenClaw installs, including persona, instructions, memory and model. You re-enter your API keys after importing.
Can StarNet run local models?
Yes. Install Ollama, then pick Ollama as the provider in StarNet. The README gives qwen3:8b as an example model.
Is StarNet safe to give full access?
Start with "ask for approval" rather than "full power". That way the agent checks with you before it acts. Each agent also runs with its own workspace and permissions, so you control what it can touch. Want help planning a safe agent setup? Grab a free AI strategy session.
