What is MCP, and why should an athlete care?

You check your Strava. You ask ChatGPT about your training. But they don't know each other. ChatGPT can't see today's ride. Strava can't reason about whether you should rest tomorrow.

That gap is what MCP was built to close. This post walks through what it actually is, what it means for your training, and why we think it's the most important quiet shift happening in sports software right now, all without the buzzwords.

The problem in one sentence

AI coaches today are smart about training in general and blind about you in particular. Ask "am I overtraining?" and you get a textbook answer. Ask the same question to a friend who's seen your last three months of TSS? You get something useful.

The analogy: USB-C for AI

MCP (Model Context Protocol) is a standard plug that lets any AI client safely read the data you authorize, nothing more, nothing less.

Before USB-C, every device had its own charger. Before MCP, every AI app had to build its own custom integration with every data source. MCP is an open protocol (published by Anthropic, now adopted broadly) that lets any compatible AI talk to any compatible data service. One plug, many connectors.

Atomic Metrix provides the connector for your training data. Claude, Codex, and any other MCP-compatible client can plug in, with your permission, and suddenly they're reasoning about your CTL, your power curve, your plan, not a generic athlete's.

See it for yourself

Here's the same question asked to two AIs. The left one is going it alone. The right one has MCP connected to your Atomic Metrix account. Toggle it and watch the difference.

What's actually happening under the hood

It's three pieces talking to each other, and your data never sprawls beyond where it should. Hover each stage to see what it does.

A few things worth noting: the AI client runs on your machine (your terminal, your editor, your browser). It doesn't get your password. It doesn't get your entire account dumped into its context. It asks for specific tools ("what's my CTL?", "what did I ride last Sunday?"), and our MCP server answers, only with what you've allowed it to answer.

You stay in control

This is the part that matters most, and the part easiest to miss. MCP isn't "connect and the AI sees everything forever." It's closer to a permission slip you can rewrite at any time. Try flipping these toggles:

In the real product, you grant scopes during OAuth signin (one browser click per scope). If an AI asks for something you didn't grant, the request is denied at the server, not at the AI. The safeguard isn't "we hope the model behaves." It's "the model literally cannot reach what you didn't authorize."

What this unlocks, in plain words

The protocol is abstract. The outcomes are concrete. Here are three things you can actually do once MCP is connected. Tap any card to see what happens behind the scenes.

Every one of these works today with Claude, Codex, or any MCP client, we have 70+ tools exposed, covering activities, power analysis, training plans, and workout scheduling.

Try it in one command

If you already have Claude or Codex installed, this wires Atomic Metrix into your AI client:

It writes one config line to your client, opens the browser for OAuth, and you're connected. If you don't have an account yet, sign up first, it's free while we're in beta.