Bounce is an iPhone app, in testing, that reads a person’s sports and health history from the apps and devices they already use, and suggests only the next one to three sessions. There is no plan to pick and no form to fill in.
It is multisport by design. Most apps are built for one sport: a running plan, a gym log, a cycling tool. In Bounce, every sport lands in one timeline and counts in one training load and one recovery picture, and one plan balances them: a run, a padel match and a strength session in the same week, with rest where the body needs it.
A round mascot does the sports. Claude does the reasoning.



Two parts. Bounce, the iPhone app, reads what the phone and the person’s apps know, keeps it, shows it, sends sessions to the Apple Watch, the calendar and Garmin, and checks every plan with a safety gate before the person sees it. The coach is Claude: for each request, a small relay sends what the phone knows to Claude Sonnet 5.5 with the coach’s rules, and Claude answers with the plan and what changed in its memory. The relay keeps nothing: no database, no logs.
The data stays on the iPhone. The sessions, the person’s words, the coach’s memory and its notes live on the phone. Each request carries what the coach needs for that run, and the phone applies the answer part by part.
Setup. One step per screen, with the mascot at each: Sign in with Apple, Apple Health, the other sources (optional, one screen), then the first plan, written in about a minute. No Claude account is needed.
An earlier version ran the coach as a Claude Code routine in the person’s own Claude account, with the data in a Google Drive folder. A run took 12 to 40 minutes and the setup had many steps, so the coach moved to the API: a plan now comes in about 30 seconds.
Sources. The phone reads:
- Apple Health, including what the Apple Watch, Garmin, WHOOP and Oura write to it: workouts, heart rate, heart rate variability, sleep, VO₂ max.
- WHOOP, Oura, Withings and Polar directly, Intervals.icu, and the Garmin account export for the years before.
- The iPhone calendars: free time in the coming week, and past sport events.
- Location and weather, everyday movement, and the photo library, where sports in photos are detected on the phone.
With the person’s sign-in, the phone also reads:
- Strava: activities, their names, who took part, heart rate zones and gear.
- Gmail: race entries, bookings, gym and club memberships, and care emails such as physio or imaging appointments. Google Calendar.
- Spotify: the person’s taste, for what to listen to during a session.
Each source on the phone has a page that says what it reads and why. Turning a source off disconnects it and deletes what it gave.
When it thinks. A message, a check-in, a correction, or new sessions read when the app opens ask the coach. What comes within 20 seconds goes in one request. Today answers at once, shows a short wait, and, when the plan lands, one calm line on what changed: “Friday’s usual intervals become rest: last week carried 1.6 times your usual load and HRV is below your normal.” In the live tests, a plan takes 11 to 62 seconds. The request goes on while the app is in the background, so opening Spotify meanwhile does not lose it.
Merging. Sources overlap and disagree. The same run can be in Garmin, Apple Health and Strava, and two devices can report different sleep. Everything is merged into one timeline before the coach plans:
- A session recorded several times counts once, across watches, phones and apps. Distance comes from the GPS device, heart rate from the record that has it, and the name and group size from Strava.
- Each metric follows one source, the one with the most data in the last four weeks or the one the person picks, never an average of several.
- An event in two calendars counts once.
- The person’s corrections always win: two sessions joined, one split, “not mine”, another sport, the distance from another source.
The phone’s own engine does the merge today. On 50 known cases, Claude merges as well as the engine (17 problems against 18 with all the cases at once, every kind of correction applied), and through the API it merged two weeks of three sources in 28 seconds. Moving the merge to Claude is the next step.
Memory. The coach’s memory lives on the phone, and goes with each request:
- An athlete model: sports, baselines, goals, events and patterns. Each fact carries its evidence and its date.
- A health ledger: every injury, illness and piece of advice, with its history. It is never shortened, and an item closes only on the person’s word or on clear new evidence.
- A watchlist of what to keep an eye on, and open questions, never asked twice.
Claude never rewrites the memory: it answers with what changed, and the phone applies it. Every plan comes from a light run that reads what is new. Once a week, a deep run rebuilds the athlete model from all the evidence, so the memory does not drift, and learns from the sessions that were done or skipped. It runs in the background after a plan lands: the person never waits for it. The evidence keeps an order: the person’s own words, then reports they shared, care emails, then devices. Newer beats older. The phone checks each part before it keeps it: a ledger that lost an item is refused, and the last good one stays.
Input. One input bar takes text, voice, photos, document scans and files: “I feel a bit sick since Sunday”, “the knee is fine again”, “I want to run with friends when it is sunny”. Speech is transcribed on the phone.


Today. One sentence, the goals, and the next one to three sessions. Sessions go in free calendar slots, in good weather for outdoor sports, at places the person uses: the gym they pay for, the club, the usual park. Each card shows the steps and their minutes, and can be sent to the Apple Watch, the calendar or, through Intervals.icu, a Garmin.
Every sport. Each sport the person does gets a short setup of a few taps: where, which gear, level, what they want, how many days a week and how long. Progress and personal bests use one model for every sport, and every session counts in the same training load and recovery picture. What only one sport needs stays specific to it. Two examples:
- Weightlifting. The setup asks where the person lifts and with what: barbell and rack, bench, dumbbells and the heaviest one at home, kettlebells, pull-up bar, bands, machines. From that, Bounce builds a movement pool out of 45 exercises across 9 patterns (squat, hinge, single leg, push and pull in two directions, carry, core). An exercise that loads an injured area is paused: a sore right shoulder removes overhead pressing until it heals. Claude plans each session from the pool with sets, reps, kilograms and rest, and moves the load up with double progression (+2.5 kg on a barbell), with a lighter week every 4 to 5 weeks and never more than the heaviest dumbbell at home. After the session, one tap logs it as planned, or the sets can be adjusted. The estimated one-rep max and the best for 5, 3 and 1 reps update by themselves. On the Apple Watch, each set is a step, with timed rests.
- Running. The setup asks about surfaces, shoes, a heart-rate strap and races. Every run counts once, even when the watch, Garmin and Strava all recorded it. Bounce keeps the best 1 km, 5 km, 10 km, half and marathon times. Suggestions come as structured sessions (warm-up, intervals at a heart-rate zone, cool-down) that go to the Apple Watch or, through Intervals.icu, to a Garmin. A low readiness day turns intervals into an easy run, and a sore Achilles removes speed work.
Cycling, swimming, tennis, padel, yoga and team sports land in the same timeline and the same weekly load, so a hard padel match on Tuesday makes Wednesday’s run easier.
Music. The coach also says what to listen to for each session, from the person’s own taste: music at a running tempo for intervals, something calm for yoga, or an episode of a show they follow for a long easy run. The session’s card opens it in Spotify.
Recovery. Every day, heart rate variability, resting heart rate, sleep and training load are compared with the person’s own normal over the last 60 days, and the day is rated “ready to push”, “good to go” or “take it easy”. The last four weeks show as one bar per day.
Health record. Each injury or illness in the coach’s health ledger has its own card: the body part, the diagnosis, since when, its status, and a recovery bar. Below it, the pain scores from the daily 0-to-10 check-in form a line over the last weeks (“pain 4 → 2 in 23 days, easing”), next to what it changes in the plan (“tennis paused”) and where it comes from (“email, physio and MRI”). Past injuries stay with their dates. An illness such as “sick since Sunday” is recorded with its start date and its fever, and the plan stays easy until it is over, with nothing hard in the 48 hours after a fever.


Questions. When the data is unclear, Claude asks instead of guessing: “In June 2024 you had an ankle X-ray. What happened?” The answer can be a sentence or a document. A report attached to an email is downloaded only after a tap.


You. Trends show gaps where there is no data and name the device behind each line. “What your coach knows” shows the coach’s memory: its open questions, the health ledger and its history, the watchlist, and the athlete model. A tap on an item shows why: its evidence, with dates. “This is wrong” sends a correction that the next request applies.
Coaching standards. The coach follows 96 standards drawn from 113 sources in the sports-science and sports-medicine literature, from the IOC, ACSM, WHO and others. They cover training load, intensity, recovery, HRV and sleep, illness (the neck check), the return from injury (pain monitoring), tapers, beginners, masters athletes, heat and travel, and red flags. Each standard gives the rule, its evidence, and how to check it in a plan.
Safety gate. Before a plan reaches the screen, the phone checks it. It refuses training through a fever the person reported, any session after chest pain or fainting until a doctor has cleared it, sessions that load an acute injury, and load spikes, such as a run more than 10% longer than the longest of the last 30 days. A refused plan leaves the last good plan in place, and the next request says why.
Tested for real. Eight journeys run live, through the app’s own code and the real relay and Claude, step by step over weeks of a person’s data: a 10 km build, a cold then a fever, a knee return, overreaching, a beginner of 52 with chest tightness, a strength plateau, a multisport week, and a triathlete with jet lag. A judge scores each plan against the standards. Three plans from the real runs:
- A cyclist has a fever of 38.4 °C, aches and a chest cough: no riding, rest, a doctor soon, and urgent care for chest pain, breathlessness or feeling faint. Short easy spins come back after 48 hours without fever.
- A runner’s knee reaches 5/10 in a run-walk and is still at 3 the next morning: one stage back, hip work and walks, then the last pain-free stage once the knee has been calm for two days.
- The beginner feels chest tightness on a hill: see a doctor this week, call 112 if it comes back, and only gentle walks until then.
Every plan judged so far passed, with scores from 3.4 to 4.0 out of 4. The 24 steps of the eight journeys run side by side in about 10 minutes. The runs also find bugs: the first safety gate read “no fever” in a ledger title as a fever. It now reads the ledger’s fever field, never the coach’s words.
Privacy. The data is on the iPhone. A request to the coach goes through the relay to Anthropic’s API, which does not train on it, and the relay keeps nothing. A request carries no exact location, no email address and no photo except the ones the person attaches, and people’s names in titles become “[name]”. One screen shows each step of the last update: what was read, what is kept, what went out, and what came back.
Bounce is in testing on TestFlight.