Guide

Can AI Be Your Cycling Coach? A Practical Guide for 2026

Joules June 23, 2026 · 9 min read

You already ask ChatGPT or Claude about your training. The advice sounds reasonable and then falls apart the moment it has to know you. Here's why a raw chatbot can't coach cycling on its own, what it actually needs, and how to set it up so it can.

Ask a chatbot "what should I ride tomorrow?" and you'll get a confident answer built on nothing. It doesn't know your fitness, your schedule, or the event you're pointing at. Ask again next week and it has forgotten everything. That's not a knowledge problem — these models know more training science than most riders. It's a memory and context problem. A coach's value isn't the textbook. It's that they remember your last eight weeks and adjust the next one.

The good news: that gap is now closable. This guide walks through what an AI coach needs to be useful, how to set one up today, and where the honest limits still are.

Why a Raw Chatbot Gives You Generic Training Advice

A standalone chat has three structural blind spots for coaching:

• No memory between sessions. Each conversation starts cold. The plan it wrote last Tuesday is gone, so it can't progress you — it can only restart you.
• No numbers, only vibes. Real coaching runs on measured load: how much stress each ride adds, how that accumulates into fitness, and how fatigue offsets it. A chatbot guessing "that sounds like a hard week" can't place intensity with any precision.
• No accountability loop. A good plan gets harder over time on purpose. Without a record of what you actually did, a chatbot has no way to tell whether to push or pull back — so plans quietly drift easier, week after week.

None of this means AI is the wrong tool. It means a chatbot needs the same things a human coach needs: a profile, a goal, a plan, and a feedback loop. Give it those and the same model that wrote you a generic answer starts writing a useful one.

What an AI Coach Actually Needs to Know

Whether you wire this up by hand or use a tool that does it for you, these are the inputs that separate real coaching from a polite guess:

1. An athlete profile

Your experience, your weekly hours, your strengths and limiters, and — most importantly — your A-race or goal event and its date. Training without a target is just exercise. A coach plans backward from the event; your AI should too.

2. Fitness, fatigue, and form

The three numbers that drive every good decision. Fitness is your accumulated training load. Fatigue is the recent cost of it. Form is the gap between them — roughly, how fresh you are. You don't need a power meter to track these; duration and perceived effort (RPE) get you a usable estimate. But the AI has to actually hold these numbers and update them, not re-improvise them each chat.

3. A season plan and a weekly structure

Good training has shape: base, build, peak, taper, then recover. Within that, each week has intent — an interval day, an endurance day, rest placed on purpose. If the AI can see the season and the current week, it can answer "what should I ride tomorrow?" in the context of where you are, not in a vacuum.

4. A way to log what really happened

This is the loop that makes it a coach instead of a planner. "Did 60 minutes, felt like a 7 out of 10." From that, the system derives the load, grades the week, and adjusts what comes next. Life happens — you travel, you get sick, you ride outside with friends instead of doing the intervals. A real coach absorbs that and resolves the week forward. So should your AI.

How to Set This Up Today

There are three honest paths, from most manual to most automatic.

Option A: Do it by hand with a long-running document

Keep a single training document — profile, goal, current plan, a log of every session — and paste it into the chat each time. It works, and it's free. The cost is discipline: you're the memory now. You have to maintain the file, recompute your load, and re-feed context every session. Most people quit this within a few weeks, which is exactly why the plan drifts.

Option B: Connect a read-only data source over MCP

In 2026, the big shift is MCP — the Model Context Protocol, a standard way to give Claude, ChatGPT, and other assistants live access to tools and data. In June 2026, Strava shipped an official MCP connector for Claude. Add it and you can ask Claude real questions about your ride history — it's genuinely useful for analysis.

The catch is in the name: read-only. It can read your past. It can't build a season, prescribe a week, or hold you accountable. For analysis, it's great. For coaching, it's half the loop.

Option C: Connect a coaching tool over MCP

The other half is a connector that gives the assistant actual coaching tools — an interview to build your profile, a season planner, weekly structure, and a place to log sessions so the plan progresses. This is the category Joules is built for: you add it once, and the chatbot you already use can plan forward and keep the plan honest.

The two compose neatly. Read-only history tells the AI what you did; a coaching connector tells it what to do next. Run both in the same session and you get memory and a coach's brain in one place.

One thing worth stating plainly: on Claude, custom MCP connectors work on every plan — including the free tier (free Claude allows one custom connector). On ChatGPT they need a paid plan (Plus/Pro). So on Claude there's no paid-AI gate at all.

What It Looks Like When It Works

Here's the difference context makes. Same question, same model — one without coaching context, one with.

The second answer isn't smarter. It's informed. It knows the block, the form, the target, and the week's intent — and it closes the loop by asking for the log. That's coaching.

Want to see a whole week of this, not just one day? We wrote up a real Joules training week — day by day, plus how the coach grades it against your actual rides and adjusts the next week. No account, no connector, nothing to install to read it.

The Honest Limits

AI coaching is real now, but it isn't magic, and a few things are worth knowing before you lean on it:

• Estimates are estimates. Without a power meter, load is derived from duration and effort. That's plenty to train well, but it's a model, not a measurement.
• It still needs your honesty. Garbage in, garbage out. If you log every ride as a 5/10, the plan can't read you correctly.
• It's not your physio. AI can place intensity and recovery sensibly. It can't diagnose an injury. Pain is a stop sign, not a data point to push through.
• You need a connector-capable assistant. As above — free on Claude (one custom connector), a paid plan (Plus/Pro) on ChatGPT.

The Takeaway

A chatbot already knows the training science. What it's missing is you — your numbers, your plan, and a memory that survives past today's chat. Close that gap, by hand or with a connector, and the AI you already pay for can do something it couldn't last year: coach you, week by week, toward an actual goal.