How Coachh works

What happens when you send a message?

Every message goes through the same steps, whether it’s “2 eggs and toast”, “185 for 6” or “slept badly, back is tight”:

1. A fixed safety screen runs first, before any model. Words that point to a medical emergency, a mental health crisis or an eating disorder get a direct answer and real help instead of coaching (see what the coach won’t do).
2. The numbers are picked out. Weights, sets, meals and weigh-ins are pulled from what you typed. On an iPhone with Apple Intelligence this runs on the phone where it can, and then never leaves it; otherwise it goes to a model as below.
3. The coach writes a reply. A model gets your message with a compact summary of your recent logs and daily health totals, what Coachh remembers about you, summaries of your conversation and your progress. It can look up your earlier meals, workouts and weigh-ins to answer a question about them.
4. A fixed check reads the reply. Before you see it, a check that doesn’t use AI blocks advice to eat under a safe daily minimum, lose weight too fast, fast for days, purge or take performance or weight-loss drugs.
5. The app logs it. What was logged shows as cards in the chat, so you can see it and fix it (“actually it was 2 eggs”).

Which AI models does Coachh use?

Coachh’s replies, plans and meal estimates are written by AI models. These are the companies involved, exactly as our privacy policy names them:

Ask the coach whether you’re talking to a person and it says it’s Coachh’s AI coach, not a person. It doesn’t pretend otherwise.

Is your data kept or used for training?

Not by the model providers, for text. On every request, Coachh tells OpenRouter to use only providers that don’t train on your data, and only endpoints with zero data retention: a request may be held in memory while it’s processed, and is never stored. OpenRouter itself doesn’t store what you send or what comes back; it keeps metadata such as how many tokens a request used and how long it took.

Voice is different, and we’d rather say so. Google’s paid Gemini API doesn’t use what we send to improve its products, but keeps prompts and responses for up to about 55 days to detect abuse and meet legal requirements, then deletes them. For replies read aloud, we also ask Google not to store the request. The xAI account used for test calls has had zero data retention on since October 9, 2026.

Coachh itself keeps your chat and logs in your account so the coach can remember them. You can export or delete everything from the app; there are no ads, and your data is never sold.

What is never sent to a model?

• Progress photos. The coach never sees them; only meal photos you send are analyzed.
• Health notes you haven’t kept. When you mention a diagnosis, a medication or something your doctor said, Coachh asks “Keep this?”. Until you tap Keep, it isn’t sent to any model, and health notes are never sent to the xAI voice test at all.
• Anything you told it to forget. A forgotten fact is never sent again and is deleted 30 days later.
• Raw Apple Health samples. Heart-rate samples, sleep stages and other raw samples stay on your iPhone; only daily totals, weigh-ins, body-fat readings and workout summaries are uploaded.
• On the xAI voice test: nothing about sleep, heart rate or HRV, pain or injuries, body measurements or body fat, and none of your chat.

Settings › Memory shows everything Coachh remembers. You can edit or delete any of it, pause remembering, or forget everything.

Where do the nutrition numbers come from?

Every item you log records where its numbers came from, and the app shows it:

An item keeps the model’s own estimate, marked as one, when no USDA food matches confidently, for branded and chain-restaurant items where it uses published nutrition, or when USDA’s number for those grams is under 0.55 or over 1.8 times the model’s, which usually means the match is the wrong food.

How are the numbers checked?

• Barcode numbers are checked before they’re logged. Open Food Facts is crowd-sourced, so Coachh looks for the common mistakes: kilojoules typed as calories, calories that don’t match the protein, carbs and fat, one serving’s numbers typed as per 100 g, and numbers that can’t be right (over 900 calories, or over 105 g of protein, carbs and fat, per 100 g). Then you see “These numbers look off” and a button to scan the label.
• The shared food list needs agreement. When you scan a label for a barcode no source has, its numbers (never the photo, never who scanned it) go on Coachh’s food list for the next person. A product’s name is only shared, and USDA’s numbers only replaced, once two people’s scans agree.
• Photo estimates are bounded. They’re shown as a range, wider when the estimate is less sure, and an unsure photo gets one question rather than a confident guess.
• Estimates wait for you before Apple Health. Exact numbers (a barcode, a label, your own food) go to Apple Health straight away. An estimate goes only after you tap “Looks right” or correct it, and it’s marked as an estimate.
• Targets are bounded too. Suggested calories never go under 1,200 a day (1,500 for men), and a weight-loss deficit is at most 500 calories or 20%. The formula is on our TDEE calculator.

How accurate is it?

We tested Coachh’s meal estimates on October 10, 2026. It’s our own test, not an independent study, and these are the results as they came out:

Read the sentence result with its catch: most of those descriptions included amounts (“150 g grilled chicken breast”), and they were scored against the same USDA data Coachh uses. So it shows how well Coachh reads a meal, not how close it gets to a plate whose weights you never said. Photos are much harder for every model: a 2025 benchmark of large models on Nutrition5k photos found average calorie errors of about 33% to 73%, depending on the model. Our own target was 70% of photos within 25%, and we haven’t met it.

So when a number matters, describe the meal with amounts or scan it. Adding a note to a photo (“half the rice”) helps too.

How are programs written and adjusted?

A program is built from what you tell Coachh in setup: your goals (ranked, with one primary), the days and minutes you can train, your equipment, injuries and the style you like. A model writes a phased plan with milestones, which takes up to about a minute, and the app checks it: sessions only on days you can train, nothing your equipment or injuries rule out, and at most four lifting days for a beginner. Each session keeps the same exercises week to week while the load changes, with a planned lighter week every few weeks.

Progression. Main lifts are loaded from your estimated max and the effort the plan asks for, and never jump more than 10% over your heaviest set last time. Accessories go up when you beat the top of their rep range. Our one-rep max calculator uses the same math.

Sleep and soreness. Each morning’s check-in (said in the chat, no form) and Apple Health give a readiness score: how you slept, your energy, any pain or soreness, and your resting heart rate and HRV against your own 14-day baseline. On a “go a bit lighter” day, accessories lose a set, the main lift is 5% lighter and sets stop around RPE 8. On a “take it easy” day, everything loses a set, loads drop 10% and sets stop around RPE 7. Pain of 7 out of 10 or feeling sick always means taking it easy, and a sore area means moves that leave it alone today without changing your program for good. The plan says why it’s lighter, and “Do the full session” puts it back.

Changing the plan goes through a review. An injury, a new schedule or a real plateau after three weeks of following the plan is a reason to change it; switching goals after a week gets an honest no, with the reason.

What won’t the coach do?

Coachh is a fitness and nutrition coach, not a medical service. Its safety rules come before its personality, even when someone asks it to skip them:

• It doesn’t diagnose. It adjusts training conservatively around pain and suggests a professional when pain is persistent, getting worse or unusual.
• Emergencies stop the coaching. Chest pain, fainting, sudden weakness, trouble breathing, or bowel or bladder changes get “stop training” and urgent care or 911, and no plan for the day.
• A crisis gets a person. Thoughts of suicide or self-harm get the 988 Suicide & Crisis Lifeline (call or text 988 in the US) and no coaching in that reply. Abuse at home gets the National Domestic Violence Hotline.
• Eating-disorder patterns end the weight-loss advice. Restricting for days, purging or feeling out of control around food gets encouragement to eat something today and a pointer to a doctor or eating disorder specialist, such as ANAD’s helpline.
• No crash diets. Never under about 1,200 calories a day for women or 1,500 for men, or losing more than 1% of body weight a week, unless a clinician supervises it. No weight-loss targets for someone underweight, and no deficit or new intense training during pregnancy or breastfeeding without their doctor’s or midwife’s OK.
• No drugs advice. No dosing, cycling or sourcing for steroids, SARMs, growth hormone or prescription medication, and never laxatives, diuretics, purging or multi-day fasts. Supporting a medication your doctor prescribed is fine.
• No promises and no shame. No guaranteed results, and no “good”, “bad” or “cheat” foods.

Three layers enforce this: the fixed screen before any model, the rules in the coach’s instructions, and the fixed check on every reply. During a voice call the app also screens each line you say and shows a help card, with buttons to call 988, ANAD or 911, while the call goes on.

What are Coachh’s limits?

• It isn’t a person and can’t watch your form.
• Photo estimates are rough; see the accuracy test.
• No micronutrient tables, no GLP-1 dose tracker, and no import of a MyFitnessPal food history.
• Apple Watch set logging is new in version 1.0. Coachh launches on iPhone and Apple Watch on October 27, 2026; Android follows in November 2026.

For how we compare with other coaching apps, including the ones that are a better fit for some people, see best AI fitness coach apps. Questions about your data: privacy policy or hello@peopleintech.io.