StarNet

A local-first desktop app for running several AI agents at once, built around a pixel-art space station instead of a chat window or a terminal. Rust/Tauri shell, a local Node "sidecar" that talks to model providers and tools, and a vanilla-JS frontend that renders the station and reflects live runtime state. 641★ / 109 forks / 14 watchers (2026-09-27 snapshot). MIT for the code; the StarNet name, logo, artwork and sprites belong to the creator, Andrew Sims, and forks must rebrand.

T1 in this ladder's sense — the primary artifact, the project speaking for itself, which is a statement about provenance, not a verdict on whether the claims below have been checked.

The idea: layout as the permission model

The pitch is that the station's floor plan is the workflow, not a picture of it:

A room is a capability-scoped team — the agents placed in it share whatever that room grants and nothing else.
A hallway is an authorized handoff lane between rooms, not decoration.
A placed object is a real capability grant (a tool, a connector, a budget), not a mock-up.

That is a spatial encoding of the same problem agent-permission-model tracks from the approval-UX side: what an agent may do without asking. StarNet's answer is to make the scope a place rather than a per-action prompt or a config file — move an agent's sprite into a room and it gains that room's capabilities; nothing suggests the reverse (revoking by moving out) is handled differently from a fresh grant, and the repo was not read deep enough here to confirm the enforcement mechanics behind the metaphor.

Stated principle: the UI cannot lie about state

The README states a specific constraint as a design rule: "the interface must never assert state the harness cannot prove." Concretely, the station does not simulate revenue, completed work, model activity, or spend that didn't happen — the ledgers (spend, budget, run history) are read from disk and shown as-is rather than animated for effect. This is a narrower and more checkable claim than most "transparent AI" marketing copy, and it is the kind of claim a release gate could actually enforce; the repo names one (qa:ready, invalidated by any new commit) but nothing here confirms what it checks.

Feature surface

Bring-your-own-model — OpenRouter key, or sign in with Anthropic, OpenAI, or Google; keys live in the OS keychain, never in the frontend. Local models via Ollama at 127.0.0.1:11434, no key or account (README warns local models are slower and rougher on long tasks).
Chat integrations — Telegram, Discord, Slack, Signal, Matrix.
Night Shift — agents keep working while the user is away, inside an adjustable "leash"; every away-action is logged and reviewable. This is the same unattended-autonomy shape as claude-code-auto-mode and the desktop-workspace approach in agent-workspace-linux, applied to a multi-agent station rather than a single coding session.
Recipes, skills, cron schedules — multi-step recipes, reusable skill grants, and scheduled runs with visible output, echoing agentskills-spec's progressive-disclosure grant model but scoped to a room rather than a directory.
Task Briefs — when a request is ambiguous, the agent asks one concrete question with options over any connected channel, instead of guessing. A named answer to the "agent silently assumes and gets it wrong" failure mode.
MCP connectors — Model Context Protocol servers, plus paste-a-key or OAuth connectors.
Voice — push-to-talk dictation in, one station voice out.
Import from OpenClaw/Hermes — reads an existing openclaw or hermes-agent home directory and creates a StarNet agent from its persona, instructions, memory and model; API keys are not transferred and must be re-entered.

Where it sits next to the spoke's other local-first agent desktops

openhuman is the closest prior entrant: also a local-first Tauri desktop app for running agents, also with Ollama support, chat/voice, and positioning against openclaw and hermes-agent. The two diverge on what they're selling as the differentiator. OpenHuman's is a persistent Obsidian-style memory tree (accumulation as the core loop, closer to the gbrain lineage, cross-wiki). StarNet's is spatial capability scoping and an explicit no-fake-state contract, plus being multi-agent by default (a station of several agents in different rooms) rather than one personal assistant. Both import from the same two sources (openclaw, hermes-agent), which keeps showing up as the de facto interchange format for agent personas between these smaller independent projects — neither StarNet nor OpenHuman defines it; both just read it.

Distribution and maturity

Installers ship from a separate repo (starnet-releases) for Windows 10/11 x64, macOS Apple Silicon, and macOS Intel; Linux builds exist internally but are not a supported public target. Windows is described as the most-tested platform, and the project calls itself an early release. 10,000+ commits sit on a feat/harness-backend branch, which reads as active, fast-moving development rather than a finished product — closer to oh-my-pi's pace than to ai-marketing-claude's one-shot drop.

Recorded weaknesses

Everything above is the project describing itself; nothing here has been run, benchmarked, or reviewed independently. The room-as-capability-scope mechanism is stated as a design principle, not demonstrated with a walkthrough of what actually happens if an agent tries to act outside its room. "The interface must never assert state the harness cannot prove" is the project's own claim about itself, not something verified by reading the sidecar code. Star/fork/watcher counts and the commit count are a single 2026-09-27 snapshot.

openhuman · claude-cowork · hermes-agent · openclaw · agent-permission-model · agentskills-spec · claude-code-auto-mode · agent-workspace-linux · andrew-sims · synthesis