AgentsAutonomy & tool use
StarNet makes a pixel-art station map control AI agent permissions, but a creator review finds few business uses
StarNet, a free MIT-licensed desktop runtime, makes a pixel-art space station map into the agents' actual permissions and handoffs. A creator review finds it the most fun and polished of the office-style agent tools he has tested, yet thin on practical business uses.
AT Pulse research deskOctober 7, 2026 · 5 min read · AI-assisted analysisStarNet's main idea is that the map is the workflow. According to the project README, each room is a team with defined permissions, each hallway is an approved handoff lane, and each object placed in the station grants an agent a real capability. The visual layout is not a dashboard on top of the governance; it is the governance. [4]
The design is local-first. Users bring their own model keys, from providers such as Anthropic, OpenAI, Gemini, xAI and OpenRouter, or run local models through Ollama. Per project materials, keys sit in the OS keychain, while memory, transcripts and spend records stay on local disk. Each job has a default $2 spend cap, with optional daily, per-agent and per-run limits. [4] [7]
In the linked video, an SEO creator who promotes his own paid community sets up a Grok-backed agent. He praises the smooth onboarding and the depth of features, which include approval modes, app connectors, a skill library and agent crews. He also says the practical use cases are limited, notes a lag after sending messages, and calls the interface very layered. [1] [3]
The project calls itself early-stage. It ships releases quickly, with v0.13.1 the latest, and publishes no evaluations or task-success metrics. The maintainers say Windows gets more testing than macOS. The only launch coverage found comes from an aggregator that summarizes the repository and does no original reporting. [4] [5] [6]
StarNet builds governance into its visual layout and keeps keys, transcripts and spend data local, per project materials.
Read the full assessment
Implication: it may make multi-agent permissions easier to inspect, but no reviewed source has tested that or measured task success.
Executive brief
StarNet draws AI agents as characters in a pixel-art space station, and the layout is meant to do real work. The free MIT-licensed desktop app had 1.2k stars when reviewed. The linked post is a promotional review by an SEO creator.
Read the full section
StarNet draws AI agents as characters in a pixel-art space station, and the layout is meant to do real work. Its README says a room is a team with set permissions, a hallway is an approved handoff lane, and an object placed in the station grants a real capability. In other words, the map you draw is the workflow the agents run (GitHub: androoAGI/starnet). The free MIT-licensed desktop app had 1.2k stars when reviewed. The linked post is a promotional review by an SEO creator. He calls StarNet the most fun agent tool he has tried this year but won't run his agency's workflows on it, because he sees few practical uses so far. No independent benchmarks exist.
What changed and event timeline
StarNet hits GitHub Trending
AIToolly reports that androoAGI's local-first desktop agent runtime debuted on GitHub Trending. The article is a summary of the repo, not original reporting ().
First aggregator coverage
AIToolly describes the bring-your-own-key model, local execution and visual monitoring of agent collaboration. All of it comes from project materials ().
- Also
Early Oct: v0.13.0 and v0.13.1 ship
The latest release, v0.13.1, fixes web requests, group chat and crew movement. The releases page shows a fast pace, with roughly ten releases since early September (;).
The creator review is posted
The Reddit post and video set up an agent connected to Grok. The reviewer praises the onboarding and the depth of features but calls practical use cases thin (;).
Capabilities and access
- Version and platforms: v0.13.1, Windows 10/11 64-bit and macOS (Apple Silicon and Intel). Linux is not publicly supported (starnetos.com).
- Cost: free to use with your own keys. An optional paid "StarNet Credits" plan manages keys for you (starnetos.com).
- Model providers: OpenRouter, Anthropic, OpenAI, Gemini, xAI/Grok, Kimi, Mistral, DeepSeek, Ollama for local models, and any OpenAI-compatible endpoint (starnetos.com).
Read the full section
- Version and platforms: v0.13.1, Windows 10/11 64-bit and macOS (Apple Silicon and Intel). Linux is not publicly supported (starnetos.com).
- Cost: free to use with your own keys. An optional paid "StarNet Credits" plan manages keys for you (starnetos.com).
- Model providers: OpenRouter, Anthropic, OpenAI, Gemini, xAI/Grok, Kimi, Mistral, DeepSeek, Ollama for local models, and any OpenAI-compatible endpoint (starnetos.com).
- Features shown in the video:
- Agent personalities and a choice between "ask for approval" and "full power" (00:28)
- Selectable reasoning levels (01:12)
- An app catalog with Gmail, Notion and GitHub (04:54)
- A skill library (06:27)
- A crew of agents with set roles (08:03)
Technical analysis for researchers and developers
- Stack: the desktop shell is Tauri v2 (Rust), the UI is vanilla JavaScript, and a Node.js 18+ sidecar runs the agents.
- Where data lives: API keys go in the OS keychain and never reach the frontend.
- Isolation: each agent run has its own workspace and permissions (GitHub).
Read the full section
- Stack: the desktop shell is Tauri v2 (Rust), the UI is vanilla JavaScript, and a Node.js 18+ sidecar runs the agents. The UI talks to the sidecar over localhost using HTTP/NDJSON and SSE (GitHub).
- Where data lives: API keys go in the OS keychain and never reach the frontend. Memory, transcripts, spend records, tasks and schedules are stored on local disk (GitHub).
- Isolation: each agent run has its own workspace and permissions (GitHub).
- Spend controls: a $2 default cap per job ($50 maximum), plus optional daily, per-agent and per-run limits and a stop control (starnetos.com).
- Evaluation: the project publishes no evaluations or task-success metrics. The maintainers say Windows is tested more than macOS (GitHub).
Claims and evidence
No independent security audit or performance evaluation was found.
Read the full section
| Claim | Source | Type |
| Real model calls and tool use, not a simulation | starnetos.com | Vendor |
| Data stays local; network calls go only to your chosen providers | starnetos.com | Vendor |
| Agents can be set up in "one or two clicks" | video 01:19 | Single creator demo |
| Some apps reach the physical world, such as finding flights and posting letters | video 05:26 | Creator, not demonstrated end to end |
| GitHub Trending debut | AIToolly | Aggregator |
No independent security audit or performance evaluation was found.
Context and prior work
The reviewer calls StarNet a "more fun" version of Paperclip (video 07:34). Paperclip is an MIT-licensed Node.js/React tool that organizes agents into a company structure, with org charts, budgets, heartbeat scheduling, audit logs and approvals (Abduzeedo). StarNet's distinguishing idea is that the visual layout itself encodes the permissions and handoffs, rather than a dashboard sitting on top of them (GitHub).
Limitations, safety and contested findings
Read the full section
- Project maturity: the project calls itself early-stage. Local models are slower on long tasks and need large context windows and plenty of graphics memory (GitHub).
- Trademark limits: the MIT license covers the code only. The StarNet name, logo and artwork belong to Andrew Sims, so forks must use a different identity (GitHub).
- Reviewer's cautions: a noticeable delay after sending messages (04:16) and a lot of interface depth (04:44).
- Inconsistent model choice: the article says the reviewer picked Grok 4.7 over 4.6. The auto-captions say he prefers 4.6 to 4.7 (01:01), although the caption text is muddled and may contain errors.
- Reviewer bias: the review pushes the reviewer's paid community throughout (Reddit post).
Business and practitioner implications
- Where it fits: StarNet suits local, budget-capped experiments with several agents, where you want transcripts and spend data kept on your own machine.
- Production readiness: treat it as pre-1.0. Releases come frequently, there is no public vulnerability reporting process, and nothing has been independently checked.
- Setup advice: start with "ask for approval" mode, as the reviewer recommends, and keep the default per-job budget caps.
Read the full section
- Where it fits: StarNet suits local, budget-capped experiments with several agents, where you want transcripts and spend data kept on your own machine.
- Production readiness: treat it as pre-1.0. Releases come frequently, there is no public vulnerability reporting process, and nothing has been independently checked.
- Setup advice: start with "ask for approval" mode, as the reviewer recommends, and keep the default per-job budget caps.
- What to evaluate: whether layout-as-permission makes agent governance easier to audit than dashboard tools like Paperclip. No reviewed source tests this.
Sources
Read the full section
The source trail.
Sources (8)
Why I Love the StarNet AI Agent (But Won't Run My Business On It)
Feed excerpt only. A transcript of the video linked from this page is supplied (linked video auto captions; https://www.youtube.com/watch?v=4wFKKIlqtQM); automatic text may contain errors.
reddit.comStarNet: FREE Local AI Agent Harness
Related coverage; assess separately
reddit.comLinked video: Why I Love the StarNet AI Agent (But Won't Run My Business On It)
linked video auto captions
www.youtube.com