Sleep
Merged across sources, so a Watch and a sleep app staging the same night don’t double-count it into nine hours of nothing.
100% on-device · zero cloud · nothing uploaded
Your wearables measure your body. Hyperopt measures what it produces. Every morning, one screen: how you’re running, what yesterday cost, and the one thing to do first. All on your Mac.
Today you’re running at about 70%
HRV down 18% vs your 30-day baseline · resting HR +4 · 5h 20m sleep
Yesterday
Most of the day in the payments service — retry logic in charge_worker.rb. Working by 15:40.
Still open
Start here
Finish the two comments on #412 before 11:00 — Ana’s question depends on it landing. Don’t start anything new today: on your last 6 mornings like this one, you opened an average of 3 new threads and finished none of them.
The whole product
Measure the body. Measure the work. Put them in the same sentence every morning. Never let any of it off the machine.
Every wearable you own, through Apple Health.
Your whole working day, measured without timers.
The two streams meet, and become one decision.
Not a policy you trust. A design you can verify.
Biohacking
You bought the ring, the watch, the sensor. They measure beautifully and then stop at the edge of your day. Hyperopt picks up every one of those signals through Apple Health — one integration, no vendor accounts — and carries them into the only place they've never been: your work.
Oura, Whoop, Apple Watch, Garmin, Withings, Eight Sleep, Libre, Dexcom — if it writes to Apple Health, it works. No vendor accounts, no OAuth, no third-party cloud, and nothing to break when a vendor changes its API.
173
health types read
0
vendor accounts
0
servers involved
Merged across sources, so a Watch and a sleep app staging the same night don’t double-count it into nine hours of nothing.
Against your own 30-day median, not a population band. Two vitals have to agree before anything is called unusual.
Any CGM that writes to Apple Health. Meals land in the same store on the same device, so one sensor and one meal is a complete measurement.
Steps, activity segments, and logged sessions — used to check whether the day you planned is the day your body had.
From your AirPods, after a calibration you perform. The question isn’t whether you slouch — it’s whether a short night makes you sink lower, earlier.
One optional check-in, or Apple’s State of Mind. Nothing here infers your mood from your face, your voice, or your typing. That guess would be worth less than the question.
Glucose
Most self-tracking asks you to wait months for a pattern. A meal and a sensor is a complete measurement in two hours — a choice you made, a number that followed, on the same device. Hyperopt is the place both of them already live.
FreeStyle Libre 3, Dexcom G6 and G7, Levels, Nutrisense. Nothing to connect and no vendor account — pair the iPhone companion and readings arrive with everything else.
The fifteen minutes before you ate become the baseline. Then peak, time to peak, and the area under the two hours after. Not a guess from a food database — your curve.
The number that pairs with sleep. It only appears in the brief when last night moved further than your nights normally move — and it arrives with how many nights that judgement stands on.
Hyperopt is not a medical device. It has no low or high alerts and says nothing about the reading you are at right now — only what past meals did. For anything that matters, use your CGM’s own app.
What your meals did
Oats & berries
08:10 · peak at 46 min
Sourdough sandwich
12:40 · peak at 51 min
Mean 104 · variability 17% · in range 92% · 281 readings
The reading count is never hidden. A CGM day is 288 readings, and a day with sixty of them is a day the sensor fell off — not a day with unusual glucose.
The iPhone companion
The companion reads your biomarkers — sleep, HRV, resting heart rate, activity, workouts, and continuous glucose — from Apple Health. Oura, Whoop, Apple Watch, Garmin, Withings, Eight Sleep, Libre, Dexcom: they all just work. No vendor accounts, no OAuth, no third-party cloud.
The brief’s best moment is before you sit down. Read it on your phone, in bed, next to your recovery score. The Glucose tab reads straight from the phone’s own Health data, so your curve and your meal markers are there whether or not your Mac is awake.
Built for wearable owners. No ring or watch yet? The brief still works from your screen data alone.
Phone and Mac pair with a PIN and sync encrypted, directly over your own local network — the data never goes through anyone’s cloud, ours included. Syncing state between two Macs uses your own iCloud Drive. There is no Hyperopt server.
Digital productivity
Hyperopt reads your screen all day, on the machine, and turns it into hours, projects, meetings and loose ends. Nothing to start, nothing to tag, nothing to reconstruct on Friday afternoon — and when it does speak up during the day, it's for a threshold you set yourself.
The data
No tracker has ever seen this much of your actual work. Nine streams, captured continuously, correlated on-device with your biomarkers. That density is why the brief can be specific — and why the patterns are about you, not a population average.
Every word you read, write, review, or abandon — read on the Neural Engine at each app switch, plus a slow heartbeat in between.
Each app and window change, timestamped. Real switches, not app-category buckets.
The docs you edited, the branches and PRs you touched, the build errors you hit.
What you copied and where it went. Concealed entries never captured.
Every call transcribed and summarized on-device — Zoom, Teams, Meet, any app. Action items included.
Home, office, travel — from your phone, mapped offline. Where you do your best work.
Sleep, HRV, resting HR, activity, workouts — and continuous glucose. Every wearable that writes to Apple Health.
How far forward your head sits, from your AirPods. Optional, and it records how long it actually watched.
One optional check-in a day: focus, stress, mood. The label on all that data.
45s
heartbeat, plus every app switch
~250 MB
per day — text & embeddings, not video
19
built-in connectors
0 bytes
uploaded, ever
Capture on its own is just exhaust. These four turn it into hours, projects, and loose ends — with no logging, tagging, or grooming asked of you.
No timers, no start button, no end-of-week reconstruction. Time away from the keyboard is trimmed off rather than counted, so the number never flatters you — and whatever couldn’t be attributed is shown as exactly that, instead of being quietly folded into something that looks tidier.
A repo root, a set of files that keep appearing together in the same sitting, a name the on-device model proposes — and it only sticks when a second signal agrees. Your own rules override all of it, and every assignment keeps the evidence that produced it.
The pull request you half-answered, the doc that stops mid-sentence, the Linear ticket you opened four times last week and haven’t touched since. Found by being picked up repeatedly and then dropped — not by you remembering to write it down.
Come back after a real gap and you get one instant question with the likely answers already ranked — from the clock, the length of the gap, and what was on screen when you disappeared. Vanished out of a call? “Meeting ran over” is already there.
Meetings
Meeting-notes tools upload your calls to their cloud. Hyperopt transcribes and summarizes on your Mac with local models — the audio never leaves the machine, and other people’s voices never touch a server.
Zoom, Teams, Meet, Webex, Slack, Discord, FaceTime — native, in the browser, even inside Citrix or VMware. Hyperopt detects the call from the audio itself. Nothing joins your meeting.
Recording uses about 1% of one core. Transcription waits until the call ends, then runs faster than real time in the background.
Your side and theirs are recorded separately, so attribution is fact, not a model’s guess.
What you promised in the meeting shows up in tomorrow’s brief. Nothing agreed on a call gets lost.
Design review — rollout plan
Google Meet · 42 min · 2 sides transcribed
Agreed to ship the rollout behind a flag on Thursday. Ana owns the migration window; the design doc’s open question on retries is resolved — exponential backoff, capped at 5.
Your action items
During the day
The brief is the product. Everything here is off by default or needs a number only you can supply, because a threshold that isn’t yours has no standing to interrupt you — and the first false positive is the one that gets a feature switched off for good.
High navigation, low production: switching at twice your own median, dwelling under 25 seconds, typing well below your usual — all of it holding for about 25 minutes. Editor → browser → terminal is debugging as much as it is spinning, and the typing is what tells them apart.
A block classifies nothing. It takes what you were already doing, calls that the work, and puts a speed bump in front of leaving it. Slack isn’t blocked because Slack is bad — it’s bumped because it isn’t what this block is for. Every block ends on one click.
Set a daily limit on distracting time and it says something at 75%, at 95%, and when it’s gone. “Distracting” isn’t a new judgement — it’s the classifier’s own score, changeable per app. The menu bar icon warms as the budget goes, so you can check without being told.
From your AirPods, after a calibration you perform. The question isn’t “are you slouching” — it’s whether a short night makes you sink lower, earlier. The reminder is optional, off by default, and separate from the measurement.
Every one of those four passes the same gate first. It holds during a call, a screen share, a Focus, a calendar event, a fullscreen app, anything playing audio, and the first minutes back at your desk. Nothing is cancelled — the next check after you’re free says it.
Settings shows you, live, what the gate currently sees and why. A tool that goes quiet without saying why is indistinguishable from a broken one.
Morning & weekly brief
This is the product. At first unlock: how you're running, what yesterday actually was, what's still open, and the one thing to start with — chosen with your recovery in mind. On Sunday, the same honesty stretched over seven days.
The Morning Brief
No dashboard. No feed. No chat. A finished brief at first unlock — read in ninety seconds, then out of your way. Everything the app does all day exists to make these four blocks right.
Your biomarkers against your own 30-day baseline. Never a leaderboard. It comes first because the same day reads differently at 70% than at 100%.
Not “3h in Development.” Which service, which file, what shipped, what stalled. Time attribution you can check — and checkable is what earns trust.
The half-answered review. The unsent reply. The doc that stops mid-sentence. Surfaced from your screen, so nothing depends on your memory.
One recommendation, not a list. A decision made for you, tuned to how you’re running — with the evidence count attached: “on your last 6 mornings like this one…”
The weekly brief
Where the hours went by kind of work and by project. The longest unbroken stretch you managed on one thing. The threads that actually cleared the bar, next to the week your body had.
And the denominator is real: unclassified time is printed as unclassified. A weekly review that adds up to a suspiciously round 100% is one you stop believing by the third week.
Your week
Mon 20 – Sun 26 Jul2h 40m
deepest single stretch
6 of 9
threads cleared
4
projects touched
6h 20m
median sleep · 3h debt
Nothing shown until eight weeks of paired biomarker-and-work data exist. A pattern you can’t trust is worse than no pattern.
After week 8
On days your HRV runs 15% below baseline, you switch context 2.4× more often and finish about half as much.
Observed across 11 days · your data only
Patterns
A few weeks of n=1 data is noise — and you’d catch us shipping it. So the rules are strict:
The wait is displayed while it accumulates. Credibility is built, not claimed.
Marginal correlations are suppressed, not padded onto a dashboard. When in doubt: nothing.
“On your last 6 days like this one…” — always with the count. Never a bare confidence claim, never a diagnosis.
Rate focus, stress, and mood with keys 1–5 — or don’t. It labels your data; it isn’t homework.
Private by architecture
Hyperopt sees your screen, your calls and your body. That is only acceptable if none of it can leave — so none of it can. Not a promise in a policy document: a native Mac app with on-device models, no servers, and a network list short enough to print in full.
Apple’s on-device foundation models write the brief, Vision reads the screen, Speech transcribes your calls. No API keys, no tokens, no per-seat inference bill.
A real macOS menu-bar app — no Electron, no dock icon, no window stealing focus — plus an iPhone companion that carries your health data over your own local network.
On a plane, on a locked-down corporate network, in a lab. Nothing you do in Hyperopt waits on a round trip, because nothing you do needs one.
Everything is stored on your Mac and exportable. Cancel and capture stops — but nothing is deleted, and your history stays yours.
Not by policy — there is no code path that would send it.
The complete list. We’d rather print it than say “zero” and be caught rounding.
Ask it anything
Hyperopt ships a Model Context Protocol server: eleven read-only tools over the data already on your Mac. Paste one line into Claude or any MCP client and ask in plain language — no export, no upload, no copy of your life on someone else’s disk.
It answers from your day summaries, time by app and project, raw activity spans, health and glucose, meetings, open loops and churn episodes. Patterns are different: that tool returns only findings that already cleared the eight-week evidence bar. The raw numbers are all there, but a correlation the app hasn’t confirmed is never handed over as though it were one.
{
"mcpServers": {
"hyperopt": {
"command": "/Applications/Hyperopt.app/
Contents/Resources/hyperopt-mcp"
}
}
}The server opens the database with SQLite’s own read-only flag. It cannot write, delete, or trigger anything — enforced by the operating system, not promised by a list of tools.
Standard input and output, to a process you launched. No socket, no port, and deliberately no “private relay for remote access” — the nearest competitor sells one, and refusing it is the point.
What you did and what was on your screen are different questions. The second has its own switch, and while it’s off the tool isn’t present-and-refusing — it’s absent from the server entirely.
Off by default. Whatever you point at it is a model that may not be running on this Mac — that choice is yours, and this is the only place in Hyperopt where it exists.
The landscape
Nearly all of these now run on your device — that argument is over, and it was never the real one. What separates them is what they can actually see. Wearable coaches measure your body and guess at your day. Health-data scorers read the same wearable and infer your day from your calendar. Local time trackers measure your work beautifully and know nothing about your body. Screen-memory tools record everything and hand you a search box. Only Hyperopt measuresboth halves — and measures the work rather than asking you about it.
| Hyperopt | Wearable coaches | Health-data scorers | Local time trackers | Screen memory | |
|---|---|---|---|---|---|
| What it measures | Your body and your actual work output, connected | Your body only — the day is a black box | Health data vs. shallow app categories | Your work only — no physiology at all | Everything on screen, into a searchable archive |
| What you get each morning | A finished brief: state, yesterday, open loops, one next step | A readiness score and “take it easy today” | Dashboards and correlations to browse | A timesheet — built to answer “what do I bill?” | A search box — you do the asking |
| Work context depth | Semantic: which file, what shipped, what stalled mid-sentence | None | Inferred from your calendar — meetings booked, not work done | Apps, sites and documents — accurate, but no judgment | Deep, but raw — retrieval, not judgment |
| Glucose & meal response | Any CGM in Apple Health, joined to the meals you logged | Some read glucose; none see your work | Charts it beside everything else | No | No |
| Where your data lives | On your devices; sync only via your own iCloud | Vendor cloud, by design | The newer ones are on-device too — that part is settled | Local — the good ones, anyway | Local (varies by product) |
Wearable coaches: Oura Advisor, Whoop Coach. Health-data scorers: Metrya, Capacity Gauge, Longevity Arc. Local time trackers: Chronoid, Timing, Rize. Screen memory: Dayflow, Screenpipe. Comparison reflects publicly available information as of mid-2026; every product here is a reasonable choice for someone — this table is about who each one is for. Hyperopt complements your wearable; it doesn’t replace it.
Pricing
Seven days free on either subscription — cancel before day seven and nothing is charged. Lifetime is a single payment, active straight away.
Seven days free, then $29.99. Leave whenever — no notice, no email to write.
Start free trialSeven days free, then $199 a year. Two months free against monthly.
Start free trialEvery plan includes
Requires macOS 26 (Tahoe) on Apple Silicon. Mac app is a direct download; the iPhone companion is on the App Store.
FAQ
Hyperopt is built for people who already track their biomarkers — anything that writes to Apple Health: Oura, Whoop, Apple Watch, Garmin, Withings, Eight Sleep. That's where the state block and the biomarker-to-work patterns come from. No wearable? The brief still works from your screen data alone.
Your data, never. Capture, OCR, and the AI all run locally on Apple Silicon, and there is nowhere to upload any of it to — no analytics, no telemetry, no crash reports. Phone and Mac sync through your own iCloud Drive. The app makes exactly two network requests, and neither carries anything about how you work: a daily check for a new version, and a check near your renewal date that your subscription is still running. Both are named in Settings, and Hyperopt keeps working if either fails.
Only a subscription. Checkout takes an email and a card at Stripe — that is the entire account. No password, no profile, and nothing about how you work is ever attached to it. If you would rather not hand over your real address, iCloud Hide My Email works fine.
Because a few weeks of one person's data is noise, and shipping noise makes the brief a horoscope. Hyperopt waits for eight weeks of paired biomarker-and-work days, shows only large stable effects, phrases them as observations, and always shows the number of days behind a claim.
No. Hyperopt is not a medical device and makes no diagnoses, treatment claims, or health recommendations. This matters most for glucose: there are no low or high alerts, and nothing is ever said about the reading you are at right now — only what past meals did, after the fact. A missed hypo is not a harm this app will ever be in a position to cause. For anything that matters, use your CGM's own app and talk to a clinician.
Any whose app writes to Apple Health — FreeStyle Libre 3, Dexcom G6 and G7, Levels, Nutrisense, and the aggregators. There is nothing to connect: pair the iPhone companion, allow Health access, and readings arrive with your sleep and HRV. Libre 2 does not write to Health on its own and needs a bridge app. Hyperopt never asks for your CGM vendor account and never talks to their servers.
The fifteen minutes before you eat become the baseline, then Hyperopt measures the peak, how long it took, and the area under the two hours that follow. Eat the same thing four times and you have a personal response to it — same person, same sensor, same food. Meals that overlap another are kept but never averaged in, because the attribution would be wrong. Every daily number carries its reading count, since a CGM day is 288 readings and a day with sixty is a day the sensor fell off.
Eleven read-only tools: day summaries, time by app and project, activity spans, health and glucose, meetings, open loops, churn episodes, and confirmed patterns. It's off by default. The server opens the database with SQLite's read-only flag, so it cannot write or trigger anything, and it speaks only over standard input and output to a process you launched — there is no network mode, no port, and no relay. Searching captured screen text is a separate switch, and while it's off that tool isn't in the server at all. Your excluded apps and retention window apply to every tool.
No. Set capture hours and nothing is recorded outside them — not on an app switch, not on a heartbeat. An end hour earlier than the start makes an overnight window. That's on top of everything else: idle, locked, excluded apps and private browser windows are already skipped, and you can pause from the menu bar or delete any range permanently.
AirPods Pro, Max or 3rd generation, and a fifteen-second calibration you perform yourself — nothing is measured against a neutral the app guessed at. It reads orientation only: never audio, never location. It's off by default, and the reminder is a separate switch from the measurement. When the AirPods aren't in, it records nothing rather than recording good posture, and every day stores how long the sensor actually watched so two days are only compared when both were worn.
Hyperopt detects a live call from the audio system itself — any app, any browser, even virtual desktops — records both sides locally, and transcribes after the call with on-device models. Audio and transcripts never leave your Mac, and microphone access has its own permission you grant separately. Recording laws vary by region: in many places you must inform participants. That's on you — Hyperopt gives you the switch, you decide when it's on.
Password managers, private browser windows, and concealed clipboard entries are never captured. Exclude any app or website yourself — the list is shown before any permission is asked. Pause capture from the menu bar. Delete any time range permanently.
3–8% battery per hour of active use, throttled further on battery; OCR runs on the Neural Engine. About 250 MB of storage per workday — text and embeddings, not video. You control retention.
The App Store sandbox forbids the screen-capture access Hyperopt needs. So the Mac app is a notarized direct download with secure auto-updates. The iPhone companion does no capture and lives on the App Store normally.
macOS 26 (Tahoe) or later on Apple Silicon (M1 or newer) — Hyperopt runs on-device Apple Intelligence models. The companion needs iOS 18 or later. Apple's macOS 27 Siri features require M3+; Hyperopt's models don't.
Seven days free, cancel in a click. If the brief doesn’t earn its ninety seconds a morning, delete it — everything it learned goes with it, off your Mac and out of existence.