Training guides

How to use Claude as your cycling coach

Strava built its connector for Claude and for nobody else. Since June 2026 a Strava subscriber can let Claude read their rides directly, and Strava's API policy does not allow any other AI app to do the same.

We build RestOrTrain, an app that plans and adjusts cycling training with AI. We cannot read Strava either (why). This is a route we do not offer. The rider brief, the export routes and the full list of prompts are in the ChatGPT guide and the prompts page, and everything there works in Claude. This page covers what is different, and we checked it against the connector on one of our own Strava accounts on September 16, 2026.

What the connector hands Claude

You add Strava from Claude's connector settings and approve read access on Strava's side; you can revoke it from either end. The connector exposes ten tools. Three matter. The first lists activities by date range, a hundred at a time, with distance, moving time, elevation and cadence per row but no power or heart rate. The second opens one activity and returns its laps, its power bests from five seconds to forty-five minutes and its personal records. The third reads one activity's streams: time, heart rate, power, cadence, speed, altitude, GPS and a moving flag, at full one-second resolution unless you ask for fewer samples. The rest: your profile, your zones and FTP with a flag saying whether Strava estimated them, your gear, your clubs and their events, an eligibility check, a health check, and a training-plan call that returns a link to Runna, a running app. Strava's own help page has the short version.

It is read only. Nothing watches for new rides, and nothing writes to your calendar or your head unit. Claude looks at Strava when you ask and not before.

Strava holds no sleep, HRV or resting heart rate, so the connector cannot pass them on. The morning question gets answered without the numbers that matter most for it. The zones come with their source. Read it: a power zone built from an FTP you set, a heart-rate zone from a maximum, a run zone from a predicted race time. Scott Dunlap, who tested the connector on his own trail-running history in June 2026, found Strava's formula-derived heart-rate zones some way off his lab-tested thresholds. He liked it anyway, because it was fast: sorting through hundreds of activities took seconds instead of an afternoon.

One thing to know before the prompts. Claude's analysis tool, the part that runs code, does not see the connector's output directly. Claude has to copy the numbers it pulled into the code it writes. Copying a thousand rows that way brings back the mistakes you went to code to avoid. Keep what goes into code small: weekly totals, or one ride at a coarse resolution. And always ask for the row count before the result.

Memory is not built for this

Claude's memory, added in August 2026, is general. It keeps a few notes about you, not your training state, and it leaves health out by default. Nothing in it changes when you ride. It was not built for training, so do not expect it to know about your knee, your FTP or what you did last week. Paste the rider brief from the ChatGPT guide at the top of each chat, with last week's review under it.

Four prompts written for the connector

Every prompt on the prompts page works here once you replace "attached is" with "pull from Strava". These four fit how the connector works. They name the date range, and they make Claude say what it fetched before it computes anything. The code step stays small.

List my activities from 1 August with the date range set, a hundred at a time, and keep paging until there are no more. Keep only Ride and VirtualRide, and tell me how many you kept and how many you dropped. Then in the analysis tool, from those rows only: hours per week starting Monday, which weeks were above 7 hours, and the longest ride each week with its date. Print the row count and the table before any comment.

List the last seven days and name the ride you think I mean by Saturday's ride. Then read its streams for time, power, heart rate and moving, at about 3,000 samples, and confirm the four arrays are the same length before you compute anything; if one is short, say so and stop. In the analysis tool: drop the first 20 minutes, find the longest stretch where I was moving and 30-second power stays within 10 percent of its median, split it in half, and give me average power, average heart rate and power per beat for each half, with the drift as the percent change. My FTP is 265 and my threshold heart rate 168. One thing to change next time. No praise.

Read my zones. Quote my FTP, whether Strava estimated it, and the source of each zone set, word for word, before you interpret anything. Then tell me which of the zone sets Strava worked out itself rather than ones I set, and what that should change about any zone-based judgement you make this season.

I missed Tuesday because of work and slept five hours last night. Tomorrow I have forty minutes, Sunday I have two hours and it is the club ride. Pull my last three weeks with the date range set, rides only, and check my club's events for Sunday's start time rather than assuming it. Before you propose anything, state the rule you apply to a missed session plus a five-hour night. Then repair the week I paste below: keep its purpose, move what you must, drop the rest, and list what moved and why. Do not add sessions to make up the lost ones.

Where it stops

Everything that stops ChatGPT stops Claude too. It only reads when asked, so the Sunday habit becomes a question instead of a paste. It is still a habit. Do not expect it to argue with you unprompted; make it state its rule before its verdict. And like every language model, it estimates a season's totals instead of computing them unless you send it to the analysis tool, and then the copying problem above applies. Nothing it writes reaches your bike computer. The route is the same as for ChatGPT: a workout in Intervals.icu's text format, which Intervals.icu then sends on.

Two things are worse here. Sleep and HRV never come through, because Strava never had them. And the connector exists for Strava alone. Rides recorded on a Garmin, Wahoo or COROS usually sync to Strava as well, which is fine. Rides that stay on the device mean exporting files, which is the ChatGPT route with a different assistant. Third parties offer bridges for Garmin and other platforms; none is built by the device makers, so treat them as you would any app that asks to log in to your account. Strava does not compute some numbers across a season, like a rolling training load or a power curve over months. Intervals.icu does, free, from your Strava or Garmin rides, and exports them as CSV.

Where RestOrTrain fits

A rider whose data lives only on Strava is better off with Claude and the connector. We cannot read Strava.

For rides that live on Garmin, Wahoo, Intervals.icu, Apple Health or the other head units and platforms on the ranking page, an app like RestOrTrain closes the gaps this page has been listing. Nothing watches Strava for Claude; in RestOrTrain the rides come in on their own. Strava never had your sleep; RestOrTrain takes sleep, HRV and resting heart rate from Garmin, Apple Health, COROS or Intervals.icu, so the morning question gets its numbers. The connector gives Claude one ride's bests at a time and leaves it to copy rows into code. RestOrTrain computes the rolling training load, the power curve over the season and your personal bests in code before the model reads a number. Claude's memory keeps notes about you; RestOrTrain remembers what you told it about your training. And where the connector is read only, this writes back: when you say the week broke, it proposes a new week, and once you confirm, the sessions go to your Garmin, Wahoo, Hammerhead, COROS, Zwift or Rouvy.

It is a phone app, on iOS and Android, with no web version and no indoor player. Beyond a few free questions a month it is $14.99 a month or $134.99 a year in the US, priced locally elsewhere. Our side-by-side with Claude and Strava goes through it row by row, and the ranking puts both next to the other apps.

Frequently asked questions

Try AI built from the ground up for athletes.

Free to start.

Published September 16, 2026