Content
# AttentionOS
> We spent billions teaching models to attend better.
> Your attention span is still about 47 seconds.
> The bottleneck isn't the model anymore — it's you.
**60 seconds to discover your attention type.** An open-source "human attention observatory":
Take a 3-stage mini-game to assess your attention, then (roadmap) use a desktop companion and local collector
to safeguard it. Data stays on your device.
**▶ Play now: https://zhonghaozhan.github.io/AttentionOS/**
| 🕹 HANDHELD (main) | 💻 TERMINAL |
|---|---|
| [](https://zhonghaozhan.github.io/AttentionOS/demo/gbc.html) | [](https://zhonghaozhan.github.io/AttentionOS/demo/terminal.html) |
| Colorful handheld RPG · Trainer cards | Dark terminal · `$ attn check-up` |
*Attention is all you need. Yours. Not the model's.*
## What it measures
The three stages correspond to Posner & Petersen's (1990) three attention networks (ANT paradigm, Fan et al. 2002),
with real-time scoring, combos, sound effects, and stage clearing:
1. **Alerting** — Simple reaction time
2. **Orienting** — Serial visual search in a rotating T-group
3. **Executive** — Eriksen flanker conflict cost
Results map to approximate percentiles from literature, producing your **attention type**
(Cheetah / Eagle Eye / Fortress / Commander / Balanced / Wanderer) + three-dimensional attributes + weekly attention prescription,
and generating downloadable trainer cards. Complete formulas, normative sources, and honest limitations:
[docs/SCIENCE.md](docs/SCIENCE.md).
Runs purely in browser: no accounts, no tracking, no network requests. Bilingual (EN/CN, switch in top-right corner).
## The ladder
| Tier | Form | Status |
|---|---|---|
| **AttentionOS mini** | 60s browser assessment (this page demo) | ✅ Playable |
| **AttentionOS** | Desktop companion + GUI Dashboard + local collector ([desktop/](desktop/)) | 🛠 Runnable: six species, edge docking, flow hiding, 25min rest invitation, five-metric report |
| **AttentionOS Pro** | CLI + MCP server + agent skill + ASCII pet (`attn pet`) | ✅ Available ([docs/MCP.md](docs/MCP.md)) |
**Your Agent can learn not to disturb you**: register [MCP server](docs/MCP.md) and
Claude Code will check `attention://state` before interrupting you — when you're in a deep focus block,
it will batch questions. Connect with [hooks](docs/AGENTS.md) and upgrade your timeline to
**human-machine joint attention**: agent work while you switch away doesn't count as distraction,
agent waiting minutes become `Agent-Wait Cost` metrics, and ASCII pets in the status bar will remind you ⏳"waiting 5m".
## `attn` — CLI prototype
Local-first attention collector (macOS):
```sh
cd cli
python3 attn.py demo # Generate one day's synthetic data, experience immediately
python3 attn.py report # htop-style attention daily report
python3 attn.py collect # Real collection (data only written to local SQLite)
```
Five open metrics (Focus Half-Life, Context Switch Rate, Interrupt Load,
Attention Budget, Recovery Debt) defined in [docs/METRICS.md](docs/METRICS.md).
## Principles
- **Local-first, forever.** Attention data stays on your device; sharing only happens when you export it.
- **No biometrics.** Only window focus and input rhythm are considered, no camera or wearable devices involved.
- **Protective, not exploitative.** This is an audit of "who's hijacking your attention",
not a productivity score for your boss.
- **Anti-burnout as first-class citizen.** Systems ignoring thermal management will burn out.
## Contributing
If you find it accurate, star this project as the best way to support open-source.
If you disagree with metric definitions, open an issue — the vocabulary itself is this project's core output.
MIT licensed.
