GoldieBench Blog · 9 min read
Startnet OS Reviews: What StarNet Is, My Verdict, And How To Run It For Free
Startnet OS reviews: it's really StarNet. The name explained, my 3.5/5 verdict, the setup, and which free local models can run it, scored on GoldieBench.
Startnet OS reviews are reviews of StarNet, a free, open-source AI agent harness, and my quick verdict is 3.5 out of 5.
"Startnet OS" is a misspelling of StarNet, and StarNet isn't an operating system, because it's a desktop app that runs AI agents inside a pixel-art space station on your own computer.
I'm Julian Goldie, and in the video above I set it up, create an agent and walk through the station.
This page explains the name, gives you my verdict and scores, shows the setup and features, and then covers the free local route with real GoldieBench numbers.
If you want my team to look at how your business can get found in Google and AI search, you can book a free SEO strategy session.
Quick verdict and scores
My quick verdict is 3.5 out of 5.
These scores are my own opinion from my own install of v0.13.1, and they aren't GoldieBench scores.
| What I scored | My score | Short reason |
|---|---|---|
| Fun and design | 10/10 | It's the most enjoyable agent interface I've used. |
| First-run setup | 8/10 | The first agent is quick, but my source install needed npm ci. |
| Brain and model choice | 8/10 | It takes subscriptions, API keys and local models. |
| Permissions and control | 9/10 | Web and file access are switches on every agent. |
| Crew and multi-agent | 9/10 | You can recruit, deploy and summon a whole team. |
| Chat speed | 6/10 | There's a slight delay after every message. |
| Practical business use | 5/10 | I don't think there's a massive amount of practical use cases yet. |
| Maturity | 6/10 | The README calls it an early release. |
| Value for money | 9/10 | It's free, and a local brain costs nothing per message. |
| Overall | 3.5/5 | It's a fun agent harness, not an operating system. |
My full scored review is in my StarNet reviews post.
Where the name "Startnet OS" comes from
The official name is StarNet, and the repo and README never call it "StarNet OS".
"Startnet" is StarNet with an extra letter, which is an easy typo to make.
The "OS" part comes from how the tool is described.
My own video on my AI News Today podcast channel is titled "NEW StarNet Agent OS is INSANE!"
I also run StarNet as a tab inside my own Agent OS, which mixes the two names together.
| What people type | What it really means |
|---|---|
| Startnet OS | It's a misspelling of StarNet. |
| StarNet OS | It's StarNet, described loosely as an agent operating system. |
| StarNet Agent OS | It's the phrase in my video title for the same tool. |
| StarNet | It's the official name in the repo and README. |
What StarNet is, and why it isn't an operating system
StarNet is made by Andrew Sims, who publishes it on GitHub as androoAGI/starnet.
The repo calls it "a living pixel-art station where real AI agents do real work."
It's open source under the MIT License and has more than 1,500 GitHub stars.
The README lists a vanilla JavaScript frontend, a local Node runtime called the sidecar and a Rust and Tauri desktop shell.
It runs on top of Windows or macOS, and you don't boot your computer into it.
It does behave like an operating system for agents, because it runs several agents at once, each with its own workspace and bounded permissions.
The README says memory, transcripts, spend records, tasks and schedules persist on disk.
The latest release on 16 October 2026 is v0.13.5.
How to set it up
The desktop app for Windows and macOS is on the StarNet releases page.
Apple Silicon Macs should use the aarch64 file, because the README says the x64 file runs under Rosetta 2.
The source route needs Node.js 18 or newer and Git, and the README starts the sidecar with node sidecar/index.js.
The README says the sidecar uses Node core modules only, but my install still needed npm ci before it would run.
The README opens the app on port 8787, and mine runs on port 8788.
The README says Windows is the most-tested platform, macOS has less real-world coverage, and Linux isn't a supported public target.
You then create an agent with a name, a pixel-art look and a personality such as warm, dry, composed or unhinged.
You pick "ask for approval" or "full power", connect a brain and choose a reasoning level from low, medium, high or max.
The agent is made in one or two clicks, and mine woke up saying, "I'm awake. Let's give the station a direction."
If you'd like my team to map out where AI agents fit your marketing, get a free SEO strategy session with us.
The features I tested
The station view zooms, changes frame and keeps everything you run in four docks you can open or drag around.
Agents level up the more you use them, a bit like a Tamagotchi for an AI agent.
The detail panel shows the model, personality, instructions and purpose, access permissions and approval prompts.
You can allow or remove web and local file access for each agent.
The Brief, Growth, History and Memories tabs show an agent's purpose, progress, past work and memory.
Chat shows status and level, and the left-hand list switches between idle and online agents and lets you manage projects and sessions.
My agent asked, "What made you want to set up an AI agent?", and I told it SEO and gave it aiprofitboardroom.com.
There's a slight delay when you send a message.
Widgets add connected apps, pinned widgets, an app chooser and a daily brief.
The app catalog, labelled "add an ability", connects Gmail, Notion and GitHub, and some apps need an API key.
It also has real-world abilities such as finding flights, sending physical letters, email and web search through Firecrawl.
The skill market has originals like ad copy testing and calendar scheduling.
The skill library has frameworks such as a decision framework for hard technical problems, making a plan, humanising content and handling a hard conversation, and you can switch each one on or off.
The Crew menu recruits roles like strategist, chief of staff, researcher and marketer, and you can deploy a role or summon a new agent.
Tasks, quests, a work list, builds and a Systems menu sit in one system.
StarNet felt like a much more fun version of Paperclip to me, and it reminded me of the first time I tested Hermes Agent.
Limits worth knowing
My honest line from the video was, "I just think it's fun to play with. I don't think there's a massive amount of practical use cases."
A viewer was frustrated that StarNet couldn't be used with Hermes Agent, and the README doesn't list Hermes as a brain.
The README says StarNet can import an existing Hermes or OpenClaw agent's persona, instructions, memory and model, with API keys re-entered.
The v0.13.5 release notes say to save a station backup first if you might ever roll back to 0.13.3 or older.
I haven't re-tested since v0.13.1, so my scores come from that version.
If you want my team to look at your site, your competitors and the keywords you're closest to winning, grab a free one-to-one strategy session.
Pros and cons
| Pros | Cons |
|---|---|
| It's the smoothest office-style agent tool I've tested. | It isn't an operating system, whatever the search term suggests. |
| The first agent takes one or two clicks. | I don't see a massive amount of practical use cases yet. |
| Every agent gets clear web and file permissions. | There's a slight delay after every chat message. |
| The crew system makes a whole team visible. | My source install needed an extra npm ci step. |
| It's free, and a local brain costs nothing per message. | Local models are slower and rougher on long tasks. |
What the latest release changes
I tested v0.13.1, so this section comes from the official v0.13.5 release notes rather than my own test.
The notes say your lead agent can now bring crew into a conversation when you ask it to.
They say delegated work gets its own threads and one live card that shows who is doing what.
They add a NEEDS YOU list that gathers every approval and question waiting on you.
They say a skill an agent creates or changes now waits for a passing evaluation and your approval.
Those changes make StarNet feel more like the agent OS people are searching for, but I'd test them before relying on them.
Who should install it
StarNet suits operators and creators who want to watch an agent crew work and learn from it.
It suits anyone who wants a free, local-first harness with clear permissions.
It doesn't suit anyone who needs a hardened agent OS for client work today.
I keep it as a tab inside my own Agent OS, which is my own product inside the AI Profit Boardroom, rather than as my main tool.
Can you run StarNet for free?
Yes, and this is the part where GoldieBench helps.
StarNet doesn't ship its own model, so a free setup means a free brain.
The README says StarNet needs no API key and no StarNet account if you use Ollama.
Its steps are to install Ollama, run ollama pull qwen3:8b and pick OLLAMA as the provider on the first-run brain screen or under SETTINGS and then PROVIDERS.
It says to pick a model that supports tools, because agents act through tool calls, and ollama show lists what a model supports.
It warns local models are slower and rougher on long tasks, and an 8B model takes about 10 GB of graphics memory while it works.
The README's example model, qwen3:8b, isn't on GoldieBench, so I can't give you a score for it.
GoldieBench scores one-shot builds out of 10, so these numbers tell you how well a model builds, not how well it chats inside a station.
The local models on GoldieBench, and which ones fit StarNet
| Local model | GoldieBench average | Scored tasks | Runner on the live page | Fits StarNet's Ollama route? |
|---|---|---|---|---|
| Qwable 5 27B Coder | 7.14 | 41 | Apple MLX only | It doesn't plug straight into StarNet's Ollama route. |
| Agents-A1 | 4.83 | 45 | Ollama, 21.0 GB | It runs through Ollama, so check that it supports tools first. |
| Qwythos 9B | 2.98 | 42 | Ollama, 5.6 GB | It runs through Ollama, but it scores low on builds. |
The strongest local builder on my board, Qwable 5 27B Coder, is MLX-only on its live page, and StarNet's README talks to Ollama.
That means the best local score doesn't map directly onto StarNet's free route.
Agents-A1 is the most interesting Ollama option, because its live page calls it an agent-tuned model built for long-horizon tool work.
It's also a 21.0 GB model, so it needs far more memory than the 10 GB the README quotes for an 8B model.
My local models leaderboard lists every on-device model I've scored.
The cheap cloud route if local feels too slow
If local models feel too slow inside StarNet, a cheap cloud brain is the next step.
I ran my own StarNet agents on Grok 4.6, because I actually prefer it to Grok 4.7.
On GoldieBench, Grok 4.7 averages 7.15 and Grok 4.6 averages 5.95, each across 20 scored tasks.
Both live pages list the same price of $2 in and $6 out per million tokens.
In my video I explained StarNet can also plug into a Claude Code subscription through the Claude CLI and into Kimi.
My StarNet reviews post lists how the bigger cloud brains score.
If you'd like my team to build your plan for getting found and recommended in Google and AI search, book a free SEO strategy session.
Also On Our Network
- Agent Operatorsthe Startnet OS setup reality, route by route
- AI Income Deskwhether Startnet OS is worth it for a business
- AI Tool Verdictis StarNet really an OS? The scored verdict
- agentos.guideStarNet compared with a full Agent OS
- aiprofitboardroom.commy StarNet AI agent walkthrough on the AI Profit Boardroom blog
If you want a clear plan for getting your content found in Google and AI search, book your free strategy session with my team.
Search for StarNet, start with a free brain and upgrade only when you need to, and that's my answer to these Startnet OS reviews.
FAQ
What is Startnet OS?
Startnet OS is a misspelling of StarNet, a free, open-source AI agent harness by Andrew Sims, published on GitHub as androoAGI. It runs AI agents inside a pixel-art space station on your own computer.
Is StarNet an operating system?
No. StarNet is a desktop app that runs on Windows or macOS. People call it an agent OS because it runs several agents with their own workspaces, permissions and memory.
Can I run StarNet for free?
Yes. StarNet is free under the MIT License, and the README shows a free route with Ollama and a local model such as qwen3:8b. It warns local models are slower on long tasks and an 8B model needs about 10 GB of graphics memory.
Which local model is best for StarNet?
The README's example, qwen3:8b, is not on GoldieBench. On the live local board, Qwable 5 27B Coder averages 7.14 but runs on Apple MLX only, while Agents-A1 averages 4.83 and runs through Ollama, so check it supports tools before using it in StarNet.
Does StarNet ship its own model?
No. You connect a brain, such as a Claude Code subscription, Kimi or Grok, an OpenRouter key, a supported Anthropic, OpenAI or Google sign-in, or a local model through Ollama.
What score does StarNet get?
Julian rates StarNet 3.5 out of 5, with top marks for fun, permissions and the crew system, and lower marks for practical business use and chat speed.
Keep reading
Hermes Herald AI Agent: Which Brain To Run It On, With Real GoldieBench Scores
GEO Examples: Four Real AI Overview Wins, The Three Moves Behind Them, And The Models Writing The Answers
Best Generative Engine Optimization Agency: Seven Picks By Region, And How The Models Behind AI Answers Score🌐 Sister-site take: read this on agentos.guide