Connect your AI

Setting it up

The connection is added from inside Claude or ChatGPT — those apps only allow a connector to be added from their own settings, so there's no one-click button on our side. It takes about a minute.

1. In Perform, open Settings → Connect your AI and copy the connection URL shown there.
2. In Claude, go to Settings → Connectors → Add custom connector. In ChatGPT, go to Settings → Connectors. Paste the URL.
3. You'll be sent back to Perform to approve the connection and choose whether the AI can only look at your data, or also take actions. Approve it, and you're done.

The Connect your AI tab has step-by-step instructions for each app, and lists every connection you've approved so you can disconnect one at any time.

Using Claude Code or your own scripts? The same tab has an “Advanced” section that issues an access key for connecting directly, without the browser approval flow.

What your AI can see

Only the communities where you're a coach or admin — never someone else's community, and never a client's private account. Within your own communities it can look at:

• Your roster — who's in each community, when they joined, their tier, and when they were last active. It can filter straight to “nobody's heard from these people in ten days.”
• Check-ins — submitted answers paired with their questions, plus who still owes you one and how overdue they are.
• Training — assigned programs and their week-by-week structure, logged sessions with adherence (scheduled vs. logged vs. skipped), and exercise-by-exercise history including estimated rep maxes.
• Nutrition — daily calorie and macro totals with a per-meal breakdown, and individual logged foods if you ask for them.
• Health metrics — body weight, sleep, resting heart rate, HRV, steps and any custom metrics, merged from manual entries, check-in answers and connected wearables using the same rules the app uses. The numbers always match what you see in Perform.
• Habits — what's assigned, what got ticked, and completion rates.
• Courses — which lessons each member has completed, and who's stalled.
• Your community feed — recent posts and their comments. Reading the feed never notifies anyone.

What your AI can do

Looking at data is the default. Actions are deliberately narrow — three of them, and they're in early access on top of the beta, so they're enabled per account on request:

• Swap exercises in a client's program. Replace one movement with another across a training block, keeping the sets, reps and notes intact. This one always shows you a preview of exactly what will change and waits for you to confirm. Applying it notifies the client that their program was updated.
• Log a workout — for yourself only.
• Log a health metric — for yourself only. Logging the same metric twice in a day replaces the earlier value.

There is deliberately no way for an AI to post to your feed, message a client, log training or nutrition on a client's behalf, delete anything, or change anyone's access. If you approve view-only access, none of the three actions above are available at all.

Example prompts

You don't need special syntax — just describe what you want. These are the jobs coaches reach for most.

Your Monday morning sweep

• “Who in my community hasn't logged anything in the last 10 days?”

  Pulls the roster filtered by inactivity, so you get a short list to act on instead of scrolling through everyone.

• “Summarise this week's completed check-ins and tell me who I should reply to first.”

  Reads every submitted check-in in one pass, then ranks them by whatever you care about — low mood, stalled weight, missed sessions.

• “Who still owes me a check-in, and how overdue are they?”

  Pending check-ins with days overdue, ready to work through.

Prepping for a client call

• “Give me a one-page brief on Sarah before our call: current program, last 4 weeks of training, her check-ins, and her weight trend.”

  Pulls from five places at once. This is the prompt that saves the most time — ten minutes of clicking in one question.

• “How has Sarah's back squat progressed over the last 3 months?”

  Best sets, estimated rep maxes, and recent history for one movement.

• “Is Sarah actually hitting her protein target? Show me the last two weeks.”

  Daily macro totals, so you can answer with data instead of a hunch.

• “Compare James's sleep and training adherence over the last month — is there a pattern?”

  Cross-referencing two datasets is exactly the kind of thing that's tedious by hand and trivial to ask for.

Programming

• “Sarah's shoulder is bothering her. Swap overhead press for landmine press in her current block.”

  You'll see a preview of every session that changes before anything is saved. Confirm, and Sarah is notified her program was updated.

• “Which of my clients skipped more than a third of their sessions last month?”

  Adherence across the roster — the conversation you want to have before someone quietly churns.

• “Find me hamstring-focused hinge variations I could program.”

  Searches the exercise library and tells you what each movement tracks.

Habits, courses and community

• “Who's stalled partway through the onboarding course?”

  Per-member lesson completion, so you can nudge the people who got stuck at lesson three.

• “What's James's habit completion rate this month, and which habit is he dropping?”

  Rates per habit, plus the day-by-day log if you ask for it.

• “What's been posted in the feed this week that I haven't replied to?”

  Posts and comments, newest first — reading them doesn't notify anyone.

Your own training

• “Log today's session: 5x5 back squat at 100kg, then 3x8 RDL at 80kg.”

  Attaches to the session already scheduled in your program where it can.

• “Log my body weight at 82.4kg.”

  Same as logging it in the app. Logging again the same day replaces the value.

Getting better answers

• Name the community if you run more than one. Otherwise the AI has to ask, or guess.
• Use first names. It looks people up on your roster — you never need an id.
• Say the time window. “Last 6 weeks” beats “recently.” Defaults are deliberately conservative, so a vague question can return less than you expected.
• Ask for the shape of the answer. “As a table,” “three bullets per client,” “just the names” — it costs you nothing and makes the reply usable.
• Follow up rather than restarting. “Now do the same for James” works, and is faster than re-explaining.
• Spot-check before acting on it. AI can misread a number. For anything that matters, the underlying data is right there in Perform.

Privacy and control

• Your data isn't copied anywhere. The AI asks Perform a question and gets an answer back, live, each time.
• Every request is checked against your account. The connection can only ever reach the communities you already coach.
• Whatever the AI reads is handled by your AI provider under their terms — worth knowing if you have client-confidentiality obligations.
• Settings → Connect your AI lists every connection you've approved. Disconnect one and it stops working immediately.

Troubleshooting

My AI says it can't find the connector

Check the URL is pasted in full, and that your Claude or ChatGPT plan supports custom connectors — it's a paid-plan feature in both.

It says it has no access to my communities

The connection needs approving on the Perform side too. Open Settings → Connect your AI and check it's listed as active. If it isn't, remove it in your AI app and add it again.

It refuses to change a client's program

Your connection is view-only. Either you approved it that way, or actions aren't enabled on your account yet — if they're not, ticking the box at approval time connects you view-only rather than failing, so it's easy to miss. Check Settings → Connect your AI: the connection shows which of the two it has. To fix it, email us to get actions enabled, then disconnect and add it again choosing the option that allows them.

The numbers don't match what I see in the app

They shouldn't differ — the same rules that power Perform's charts produce these answers. Check the date range the AI actually used first, since it may have defaulted to a shorter window than you had in mind. If it still doesn't line up, tell us.

It's guessing instead of looking things up

Ask it directly to check Perform. Being explicit — “look up her actual logged sessions” — reliably pushes it back to the data.

Support

This is a beta and we're actively shaping it around how coaches use it. If something is missing, wrong, you need access enabled, or you've found a prompt that works brilliantly, email support@prfrm.app.