MCP

The AI-Forward Cyclist's Stack: Strava MCP + Joules + Garmin

Joules June 29, 2026 · 8 min read

If you've added an MCP connector to Claude before, this is for you. 2026 is the first year you can assemble a genuinely complete cycling-coaching system out of parts you bolt onto the assistant you already use — no new app to learn, no monthly coaching fee. The trick is knowing which part does what. Here's the stack, why each piece exists, and how they compose into one loop.

The mental model: read, reason, write

Every coaching workflow, human or AI, has three jobs. It has to read what you actually did, reason about whether it's working, and write what you should do next. A human coach does all three in their head. The reason an AI assistant couldn't really coach until recently is that it could reason brilliantly but had no hands — no live read of your training, no way to commit a plan you'd be held to. MCP connectors are the hands. Once you understand the three jobs, the stack picks itself:

• Read — a data connector: Strava's MCP, a Garmin connector, or an aggregator.
• Reason — the assistant itself: Claude or ChatGPT.
• Write — Joules: the season planner, weekly structure, and the grade that drives next week.

Layer 1 — Read: get your rides into the assistant

This is the data layer, and in mid-2026 it's finally good. In June, Strava shipped an official MCP connector for Claude — add it and the assistant can read your ride history and reason over it in plain language: summarize a block, spot a fading Saturday, pull your best 20-minute power. If you ride a Garmin or Wahoo, your data already flows to Strava, so the Strava connector is usually the simplest on-ramp. Prefer to go straight to the source? There are a few Garmin routes that work in 2026 — a hosted connector, an aggregator like intervals.icu, or self-hosting.

The thing to internalize about this whole layer: it's read-only by design. Every one of these connectors brings your past in. None of them plans your future. That's not a shortcoming — it's the job. But it's exactly why the stack needs a second layer.

Layer 2 — Reason: the assistant you already pay for (or don't)

Claude and ChatGPT are the brain. They're what turns a pile of ride data into "your aerobic base is solid but you've skipped your hard days three weeks running." You don't add anything for this layer — it's the assistant itself. Worth stating plainly, because it changes who this is for: custom connectors work on every Claude plan, including the free tier — a free Claude account gets one custom-connector slot, so you can run this stack for $0 on the assistant side. On ChatGPT, using apps and connectors needs a paid plan (Plus/Pro). Either way, that's the platform's rule, not a tool's price tag.

Layer 3 — Write: the part the read connectors don't do

Here's the gap. With only a read connector, you can ask "what should I ride tomorrow?" and the assistant will give you a reasonable-sounding guess — but it's a guess, because there's no plan, no goal race, and no sense of where you are in a progression to place tomorrow against. Coaching is the write side: an interview to build your athlete profile, a season built backward from an A-race, specific weekly workouts, and a place to log what you did so the plan adapts and stays honest.

That's the category Joules is built for. It's an MCP app that gives Claude or ChatGPT a cycling coach's brain: it interviews you, builds your season, plans each week, and grades every week against your real rides — device-free, from just duration and perceived effort if that's all you've got. You add it exactly like Strava: paste the connector URL https://context.joules.cc once and sign in with your email.

How the three compose in one chat

The payoff is that you run all of this in a single conversation. The read connector tells the assistant what you did; Joules tells it what to do next, then checks whether you did it. Watch the same question get a sharper answer as the stack fills in:

That second answer is only possible because all three layers are present: the read connector supplied the real history ("you cut the intervals short"), the assistant reasoned over it, and Joules supplied the plan, the goal, and the place to log the result. Remove any one layer and the answer degrades back toward a guess.

The honest assembly order

You don't need everything on day one. The fastest path to value, in order:

• Start with the write layer. Add Joules first — it works on its own, even with no device or data connector, because it can plan from your goal and grade from what you tell it. You get a real plan in your first chat.
• Add a read connector when you want tighter feedback. Bolt on Strava's MCP (or a Garmin route) so the grading runs off your actual rides instead of self-report. This is the upgrade that makes the loop feel like a coach who's actually watching.
• Skip the rest. You don't need a power meter, a paid analytics app, or five connectors. Three layers is the whole stack.

Curious what the planning layer produces before you wire anything up? Build a sample week in your browser — pick your goal and weekly hours and watch the coach assemble it, no account or connector needed.

Want the full landscape of read connectors first? See the best AI connectors for cyclists, or the head-to-head comparison of cycling MCP servers.