Athina sits in your menu bar, reads what you are working on, and tells you in a small note when it sees a better way to do it. It asks a model with your own API key, and nothing but those calls leaves your Mac.
- Read on your Mac - the app in front, its window and the text on screen are read and kept here.
- Quiet - a cheap model picks the moments worth a look, and the strong model most often has nothing to say.
- Points at the spot - a note about one place on screen can outline it.
- Keeps your goal in mind - it can tell you when an approach will not get you where you are going.
- Your key, your cap - Anthropic, OpenAI or OpenCode, billed to you and capped at $1 an hour by default.
ON YOUR MAC ┆ AT YOUR MODEL PROVIDER, ON YOUR KEY
┆
what you are doing: the app in ┆
front, its window, the screen text ┆
│ ┆
▼ ┆
┌──────────────────────────────────┐ ┆
│ sensing │ ┆
│ nothing before Allow, while │ ┆
│ paused, or on an excluded app │ ┆
└─────────────────┬────────────────┘ ┆
▼ ┆
┌──────────────────────────────────┐ ┆
│ journal │ ┆
│ what was read, kept on this Mac │ ┆
└─────────────────┬────────────────┘ ┆
│ you switch apps ┆
│ or stop typing ┆
▼ ┆
┌──────────────────────────────────┐ ┆ ┌────────────────────────────────┐
│ triage │ ┆ │ the cheap model │
│ at most once every 20 s, and ├─┼──▶ is the latest screen's │
│ only when the screen changed ◀─┼──┤ text worth a look? │
└─────────────────┬────────────────┘ ┆ └────────────────────────────────┘
no: back to │ yes ┆
watching ▼ ┆
┌──────────────────────────────────┐ ┆ ┌────────────────────────────────┐
│ mentor │ ┆ │ the strong model │
│ 2 min since the last mentor ├─┼──▶ reads the recent screens' │
│ call, and under the hourly │ ┆ │ text, a thumbnail and what │
│ spend cap ◀─┼──┤ you seem to be after │
└─────────────────┬────────────────┘ ┆ └────────────────────────────────┘
most often │ a suggestion ┆
nothing ▼ ┆
┌──────────────────────────────────┐ ┆
│ a note under the menu bar, and │ ┆
│ a callout on the spot it means, │ ┆
│ when it means one │ ┆
└─────────────────┬────────────────┘ ┆
▼ ┆
your answer, kept in the journal: ┆
Tell Me More, Not Now (that kind ┆
of note waits an hour) or Never ┆
for This (that kind of note stops ┆
for this app) ┆
By default the cheap model is Claude Haiku 4.5 and the strong one Claude Opus 5, or GPT-6 Luna and GPT-6 Sol on OpenAI. The thumbnail can be turned off in Settings > Models, and a callout comes down as soon as its window moves or changes.
What you seem to be after is the standing understanding, a short note each mentor call rewrites. After 15 minutes of use with no mentor call, a refresh rewrites it on its own.
Every call counts against one hourly cap, $1 by default: calls slow down as the hour's spend nears it and stop at it until the next hour. Change it in Settings > Models > Spend at most, and see the mentor loop for every gate.
Athina runs on macOS 26 or later.
Download Athina-<version>.dmg from GitHub Releases and drag Athina onto Applications.
Releases are not notarized yet, so allow the first launch in System Settings > Privacy & Security (Install has the steps).
- Open Athina and choose Allow in the Setup window; nothing is sensed or sent before it.
- Grant Screen Recording and Accessibility on its permissions page (Permissions).
- Paste an API key in Settings > Models, press Save, then Test Connection (An API key).
make # list every command
make run # the app on recorded answers: no key, nothing spent
make check # lint, test and snapshots: what a change passes before a pushContributing has the setup, the rules and the docs.
- Privacy: what each call carries and what stays on your Mac.
- Setup: install, API keys, what calls cost, permissions.
- The mentor loop: every gate, tier and setting.
Athina is early, and issues are welcome. It is open source under the MIT License.