Rhythm, not content.
Ersa never reads your screen, your keystrokes, your files, nothing like that. She just reads the shape of your attention. Turns out that's enough.
The mechanics, in plain words
She watches the rhythm
How often you're switching apps. How long you actually settle into one. Whether you keep looping back to the same place. Whether the keyboard's gone quiet. All stuff macOS already exposes, nothing gets installed inside your apps, nothing reads what's on your screen.
Your actual work tools count as work: bouncing between the apps you use for work reads as flow, not chaos. It's the bouncing into everything else that she's paying attention to.
The rhythm becomes a state
A tiny on-device engine turns all that into a running read: flow (you're locked in, leave her alone), threads (attention's pulled in too many directions), overloaded (too much, all at once), or open (calm, available). It's smoothed out over minutes, not seconds, so one random app switch isn't going to set anything off.
She also pays attention to direction. If a drift's already fixing itself, she leaves it alone: interrupting someone who's already coming back on their own is worse than just staying quiet.
A state becomes a moment, sometimes
When a state sticks around long enough, and the timing's actually respectful (not mid-flow, not right after your last check-in, never during quiet mode), Ersa shows up with one small observation and exactly two choices. Different situations, different moments: scattered attention, can't-start paralysis, doom-looping on one site, a twenty-minute stall, ending up somewhere you didn't mean to be.
Most of the time, the right move is no moment at all. Her best day is the one where she stayed quiet because you didn't need her.
She learns you
Every time you respond, she learns something. Engage with a check-in and she trusts that read a bit more. Dismiss a bunch of them and she backs off. Tell her "that wasn't right, I was actually in flow," and she raises the bar before calling it drift again, specifically for you.
No quiz, no personality test, no setup screens. The calibration just comes from using it, which is the one thing an ADHD brain gives you for free anyway.
One exception, on purpose: overload never gets talked down by feedback. "I'm fine" while things are clearly escalating is a pattern she's specifically built to not fall for.
You always outrank the sensors
Tell her what's actually true ("I'm in flow," "I'm overwhelmed") and that steers her read for a while, fading out gently over time. Her guesses never override what you tell her. You're the authority on your own state, she's just the one paying attention.
You can just talk to her
Typing's the default, but there's a voice mode too. Switch it on and she'll actually say things out loud instead of just showing text, and once she's done talking, she starts listening automatically, no need to tap anything first.
She waits for a beat of silence to know you're done, then picks the conversation back up. Flip back to text mode any time, mid-conversation, whenever typing feels easier.
Held thoughts
"Hold this for me" saves it, no ifs about it, and she only confirms once it's actually saved. Then she waits, not for a timer, for an actual calm moment, because handing you a thought back mid-meltdown just gives you one more thing to drop.
She figures out urgency from how you said it. "Before they close today" comes back within hours, guaranteed. "Someday I want to think about this" waits for a genuinely quiet moment. And if you mention out loud that it's done, she just lets it go. No checkbox needed.
The privacy architecture
Conversations go to Anthropic's Claude through Ersa's own relay, tagged with nothing but a random install ID. No account, no name, no email attached. The API key never ships inside the app, and there's no dashboard anywhere for anyone to go read what you said.
Crisis language is the one exception that proves how this is built: it gets caught entirely on your Mac, answered with fixed, human-written resources, and never leaves your device at all.
The full policy is short and in plain language. Read it here.