What athletes are doing with it
We built Tempo's connector for exactly this, and it has still been something to watch. Athletes connect Tempo to the assistant they already use, Claude or ChatGPT for most of them today, and within a few conversations that assistant stops being a question box and starts behaving like a coach. It reads the last six weeks of rides before it answers. It notices that the athlete's sleep was poor on the two nights before the flat interval session and says so. It builds a twelve-week block, simulates the load, asks one round of questions, and puts the workouts on the calendar. The following week it checks compliance, sees a note that says traveling, and shifts the threshold session to Thursday without being told how.
None of that lives in a single feature. It emerges from an assistant that can read an athlete's data, apply a method, act on the plan and remember what happened last time. We have started calling this pattern agentic coaching, because "AI coach" no longer describes it. An AI coach answers. An agent reads, decides, acts and follows up.
This post is about why we think agentic coaching is where endurance training is going, what has to be true for it to work, and what Tempo has built so that it already works today.
What agentic coaching means
The first wave of AI in endurance sport was plan generators. You typed your goal, your hours and your FTP into a chat window and got back a plan. The plans were fine as far as they went, and the research on them is honest. When coaching experts rated ChatGPT-generated running plans against 22 quality criteria in a 2024 study, the plans built from a bare prompt were rated below par on most of them, and the ratings rose steadily as the prompt carried more information about the athlete. That result is not an argument against AI coaching. It is the whole argument for context.
A chat-only plan knows exactly what you paste into it. It cannot see that you actually skipped Tuesday, that your resting heart rate has drifted up for five days, that you have a gravel race in seven weeks that is already on your calendar, or that you rode the last long ride under-fuelled. It cannot put the plan anywhere useful either; every workout has to be retyped into whatever app pushes to your head unit, and by the second week the plan and reality have parted ways.
Agentic coaching closes both gaps. The agent has tools: it can fetch the rides, the wellness readings, the events, the notes and the nutrition, and it can write planned workouts, events and notes back. It has a method: a protocol for how a block is built, what the load guardrails are and when to change course. And it has memory that spans conversations, so the athlete who explained last month that they cannot train before 6 am does not explain it again.
That combination is what a good human coach does. They look before they prescribe, they follow a method, they put the work on your calendar and they remember you.
A good coach is also expensive, and not always there. Coaching runs to hundreds of dollars a month, which is why most self-coached athletes have never had one, and even a coach who answers every message does not answer at 6 am when the morning HRV reading arrives, or at 9 pm when tomorrow's session has to move. The agent is there at both hours, and its cost is a Tempo Premium subscription at $5 a month on top of an assistant that costs around $20 a month on its paid tiers, or nothing at all on Claude's free plan. That is not a claim that an agent replaces the judgement of a coach who has known you for years. It is a claim about who gets coached at all: agentic coaching brings look-before-you-prescribe discipline, on demand, to the athletes who were never going to hire one.
There is a quieter benefit too. A plan from a coach usually arrives finished; a plan from an agent arrives with its reasoning attached, in your own numbers, and you can ask why as often as you like. Why sweet spot this block and not VO2max. Why the long ride moved. Why this morning's HRV means an easy day and last Tuesday's did not. Ask a coach that every day and you are paying for their time; ask an agent and you are simply learning how training works, on your own data, at the pace you choose.
Three parts, and who owns each
It helps to be precise about what is doing what, because the vocabulary is loose and a lot of products will claim the word.
- The agent is the assistant itself: the model, its memory, and the loop in which it decides which tool to call next.
- The tools and the data are the platform: the athlete's complete training, wellness and fueling history, and the read and write operations the agent can perform on it.
- The methodology is the coaching knowledge the agent runs on: how to build a block, how to place strength work around key rides, how to taper, when to back off.
Tempo is the second and third of those. We are not trying to be the agent. Anthropic, OpenAI and Google are investing in their agents at a scale no endurance software company can match, and the assistant an athlete already uses knows things about their life, their work travel and their family calendar that a training app never will. Building a walled-off chatbot inside Tempo would give our athletes a worse agent with less memory that only sees a slice of their world.
What we can build is the platform an agent needs underneath it: the data in one model, the write access to act on it, and the coaching method written down in a form an agent will actually follow.
What Tempo has built
Tempo's hosted Model Context Protocol (MCP) server launched in April 2026. It started read-only: an assistant could see workouts, planned training, events, fitness, wellness and nutrition, including fuel plans, over a single connector with nothing to install. Since then it has been extended in three directions. The reads went deeper, with performance ceilings, per-lap data, heat metrics from core temperature sensors and a personalized reading of HRV. Write access arrived, so the assistant can create, change and remove planned workouts, events and calendar notes. And the coaching methodology itself became part of the connector: protocols for building cycling training plans and for placing strength work around them, which the agent fetches and follows before it touches the calendar. The connector page lists what it reads and writes today.
Beyond the list of capabilities, three design choices matter more than any single one.
Read before you write. Before an agent designs anything, the protocol has it pull the athlete's profile, several weeks of training load and wellness, the planned calendar and any notes on it, such as travel, illness or a race. An agent that acts without looking is a liability, and the fastest way to prevent it is to make looking the first step of the method.
One model of the athlete. Training load, sleep and HRV, carbohydrate intake and energy availability are derived from the same history and exposed through the same connector. "Why was Thursday so bad?" gets answered with the ride, the night before it and what the athlete ate, because the agent can see all three at once.
The athlete confirms. The agent simulates the projected load against guardrails, presents the plan and the projection once, and gets one confirmation before it creates anything. Afterwards it checks the result against the server's own numbers and fixes anything that breaches a guardrail. The athlete stays the decision-maker; the agent does the reading and the typing.
Bring your own agent
The most common question we get from people evaluating this is whether it locks them into one assistant. It does not, and that is the point. Tempo speaks the Model Context Protocol (MCP), an open standard for connecting assistants to tools and data, so it works with any assistant that supports it: Claude and ChatGPT are the two most athletes use today, and a desktop coding tool or a self-hosted agent connects the same way. When a better assistant appears, Tempo will work with it on the day it ships MCP support, and the athlete's method and data go with them.
It also means Tempo's coaching improves on two independent clocks. When we improve a protocol or expose a new metric, every athlete's agent gets better that afternoon. When the labs ship a stronger model, every athlete's agent gets better without Tempo shipping anything at all. A training platform with a built-in chatbot only gets the first of those.
Why the incumbents are not doing this
It is worth asking why the largest platforms in the category are not further along here, because the answer is structural rather than a matter of will.
An agent needs an API that an individual athlete can point their assistant at, with permission to write. Most platforms built their APIs for partner integrations and coaching businesses, not for an athlete's own agent. TrainingPeaks states that its API is not available for personal use, and an athlete has no way to bulk-import a plan built outside it. That is why a cottage industry of "translation layers" exists to get chat-generated plans onto calendars, and why one of those tools, TrainingDojo, describes the limit of chat-based planning as plainly as anyone: "They only know what you tell them." TrainerRoad's own case against using ChatGPT as a coach is that it "cannot directly integrate with your workout data" and cannot manage day-to-day training load. That is true of a chat window. It is exactly what a connector fixes. Every one of those limitations is a missing tool, not a missing model, and a read-only connector, however polished, still gets you a summariser rather than a coach.
There is also a business reason. A platform whose value is the plan it sells has an incentive to keep the plan-building inside its walls. Tempo's value is the model of the athlete and the method for acting on it, and both are worth more the more capable the agent using them becomes.
The market is starting to split. Strava's connector, launched in June, reads your history and cannot write. A handful of platforms, Tempo among them, let an assistant write to the calendar as well. We keep a comparison page that says plainly which platforms let an assistant read your data, which let it change your plan, and which have no official connector at all. What we have not found anywhere else is the combination: training, wellness and fueling in one model, write access to events and notes as well as workouts, and the coaching methodology itself delivered as something the agent fetches and follows, all over one open connector.
Where this goes next
Everything above is reactive: the agent acts when the athlete opens a conversation. The next step is an agent that acts when the data changes. The protocols already describe the weekly check-in the agent should run; the next step is running it without being asked. That is the direction: a morning reading that is two standard deviations suppressed, a missed key session, a note that says sick, each of which should reach the athlete's agent without waiting to be asked, with the method already loaded and the calendar already open.
The other direction is more methodology in the agent's hands. Fuel planning, heat acclimatization, return from illness and the long tail of coaching judgement that today lives in a good coach's head can be written down as protocols the agent fetches, checked against the athlete's own numbers, and improved in the open.
We are early, and so is everyone else, though few have come this far. But the shape of the thing is now visible: the coach of the next decade is an agent the athlete already trusts, running on a platform that gives it the athlete's whole picture, the power to act on it, and a method worth following. That is what Tempo is building.
How to try it
In Tempo on the web, go to Settings → AI Assistants, select the Claude or ChatGPT tab, and follow the steps to add Tempo as a custom connector at https://mcp.jointempo.ai/mcp.
Then ask for something a plan generator could not do:
"Read my last six weeks and tell me whether I have headroom under my aerobic ceiling or whether I am ceiling-limited, then propose the block that follows from that."
"Build me an eight-week block for the gravel race on my calendar. Work around the vacation note in week four, simulate the load first, and only create the workouts once I confirm."
"Check in on last week: what did I complete, how did my HRV and resting heart rate track against baseline, and what should change this week?"
Note: Tempo MCP requires a Premium subscription. Review what the assistant proposes before relying on it for your training; the protocols are built so that it asks first, and you should hold it to that.