# What is a fitness MCP server? How AI assistants read and write your training data > A fitness MCP server lets Claude and ChatGPT use your training data as tools. What MCP is, read-only vs read-write servers, security, and how to connect one. Published 2026-11-12 · Updated 2026-11-12 · revns editorial A **fitness MCP server** is a service that exposes your training data as tools an AI assistant can use, through the open **Model Context Protocol (MCP)**. Connect one to Claude or ChatGPT and you can ask "what should I do today?" and get an answer from your real data. A **read-only** server shows the assistant your history; a **read-write** server also lets it log sessions, set goals and plan your days, with your permission. **TL;DR** | Question | Short answer | | --- | --- | | What is MCP? | An open standard for plugging AI assistants into tools and data | | What does a fitness server add? | Your training, sleep, food and goals as tools | | Read-only vs read-write | Read shows history; write lets the assistant log and plan | | How to connect | Paste the server URL in your assistant's connectors, sign in, allow | ## What is the Model Context Protocol? MCP is an open protocol that standardises how applications provide context and tools to AI models. Instead of every assistant building its own integration with every service, a service publishes one **MCP server** and any assistant that speaks MCP can use it. The server describes its tools (name, purpose, inputs and outputs), and the assistant decides when to call them during a conversation. A **remote** MCP server runs on the web, so you connect it by URL and sign in, with no software to install. That is how most consumer fitness servers work. ## How is MCP different from an app integration or an API? Most fitness apps already have integrations: Strava talks to Garmin, a food app talks to a health platform. Those are fixed pipes between two products, designed by developers in advance. An API is the raw interface a developer codes against. MCP sits one level up. The server describes its tools in plain language, and the **assistant** chooses which to call, in what order, based on what you ask. "Log my run, then tell me if I am on track for my 10K" becomes a log call, a goals call and an answer, without anyone having built that exact feature. That flexibility is the point, and it is also why permissions matter: the assistant decides what to call, so the server must decide what it is allowed to do. ## What can a fitness MCP server do? It depends on the data behind it and the tools it exposes. Typical **read** tools: - Today's state: what you have logged, what is still open, a readiness or day score. - Trends over weeks: averages, bests, misses. - Goals and whether you are on pace. - History from a device platform, such as runs and rides. **Write** tools go further: log entries ("40 squats, slept 7 hours"), log meals, create something new to track, set or update a goal, write a plan for tomorrow, or save a morning brief. ## Read-only or read-write: which do you need? Several device platforms now offer official connectors. Strava, for example, launched an MCP connector on 1 June 2026 that lets subscribers' assistants read their training data; it is read-only. That is ideal for questions about past activities. A read-write server suits a different job: holding your **objective, plan and daily log** so the assistant can run the loop with you. "I did legs and slept six hours, plan me something lighter tomorrow" needs a server that can write. Many people will use both: a device connector for what a watch recorded, and a read-write server for the plan. The revns [MCP server](/mcp) is read-write, with writing behind a separate permission. ## How is a good MCP server secured? Your training data is health data, so check for: - **OAuth sign-in** with PKCE, not a pasted password or long-lived API key. - **Scoped permissions:** reading and writing granted separately. - **Short-lived access tokens** with rotating refresh tokens, and a way to revoke any connection. - **Clear write labelling:** every change the assistant makes is marked with its name, and tools that change data say so, so the assistant asks before using them. - **Encryption at rest** for the data behind the server. - **Health caution:** the server and its assistant should give training guidance, not diagnosis. ## How do I connect one? The steps vary by assistant and plan, but the shape is the same: open the connector settings, add a custom connector with the server URL, sign in to the fitness service, and approve the permissions. Step-by-step guides: - [Connect your fitness data to Claude](/mcp/claude) - [Connect your fitness data to ChatGPT](/mcp/chatgpt) - [Claude Code](/mcp/claude-code) and [Cursor](/mcp/cursor) ## What should I ask once it is connected? Start with questions where your data changes the answer: - "What should I do today, and why?" - "Plan my week. I am vegetarian and eat eggs." - "Am I on pace for my goal? What should change?" - "Log 40 squats and 20 push-ups, and that I slept six hours." - "How did my sleep look over the last two weeks?" ## What are the limits? An assistant is only as good as the data and tools behind it. It cannot see what you did not log or sync, it can misread an ambiguous request, and it is not a clinician. Keep write permission for assistants you trust, review what they log, and treat health advice as general guidance. ## What does a conversation actually look like? A typical exchange, with a read-write server connected: 1. You write: "Did 40 squats and 20 push-ups, slept six hours. Plan me a lighter day tomorrow." 2. The assistant calls a logging tool with your sentence, then a "today" tool to see your score and that sleep was under your range. 3. It calls a planning tool for tomorrow with an easier session and mobility. 4. It answers in plain language and the plan appears on your dashboard, labelled with its name. Each tool call is visible in most assistants, and tools that change data are marked so the assistant can ask you first. ## What are scopes and why do they matter? Scopes are the specific permissions you grant during sign-in. A well-designed fitness server separates them, for example: - **read:fitness** for your day, trends, goals and plans - **write:fitness** for logging, goals and planning - **read:news** for articles and guides Grant only what you need. If you want an assistant that answers questions but never changes anything, connect it without the write scope. ## Can I use more than one server at once? Yes. Assistants can hold several connectors. A common setup is a device platform's connector for recorded activity history and a read-write server for the plan and daily log. Ask questions that span both, such as "compare my Strava runs with how I slept", and the assistant calls each as needed. ## What should a server never do? - Return another user's data, or your email and tokens, through a tool. - Accept writes without validation: a value outside a sensible range should be rejected, not stored. - Hide which assistant made a change. - Present itself or its assistant as a medical service. ## How do I check what the assistant changed? Look for an activity trail. In revns, every entry, plan item and coach note written over MCP is labelled with the assistant's name, and you can disconnect any assistant from the Coach screen. Reviewing the first few days of changes is a good habit with any write-enabled connection. ## Do I need a paid AI plan? It depends on the assistant. Custom connectors are available on some plans and not others, and providers change this often. Check your assistant's current help pages for "custom connectors" or "MCP". Coding tools such as Claude Code and Cursor generally support remote MCP servers directly. The server itself, for revns, is free to connect. ## Is MCP only for coding? No. It started in developer tools, but any assistant that supports MCP can use any MCP server, from calendars and documents to fitness data. Health and fitness is a natural fit because the useful answers depend on your own data: "what should I do today" has no good generic answer, and a lot of people already ask their assistant health questions every week. ## How do I disconnect an assistant? Two places: in the assistant's own connector settings, and on the service side. For revns, open Coach and choose Disconnect next to the assistant; its access stops at once, even if the assistant still lists the connector. Reconnect any time and choose permissions again. ## Where is this heading? Expect more official connectors from device and platform makers, more assistants that support remote MCP servers out of the box, and more read-write tools as services get comfortable with permissions and audit trails. The practical question for a user stays the same: which assistant do you trust to act on your health data, with which permissions, and can you see and undo what it did? Start small: connect read-only first, ask a week of questions, then add write access once you trust the answers. ## Related - [The revns MCP server](/mcp) - [AI fitness coach and week planner](/features/ai-fitness-coach) - [Strava MCP connector vs a read-write server](/compare/strava-mcp-connector) This is general information about software, and any training advice from a connected assistant is guidance, not medical advice. ## Sources - [Model Context Protocol specification](https://modelcontextprotocol.io/specification/2025-06-18) - [Strava Support: Strava MCP connector](https://support.strava.com/hc/en-us/articles/15401531-strava-mcp-connector) ## Questions ### What does MCP stand for? Model Context Protocol, an open standard for connecting AI assistants to outside tools and data. An assistant that supports MCP can use any compatible server. ### What is the difference between a read-only and a read-write fitness MCP server? A read-only server lets the assistant see your data, for example past runs. A read-write server also lets it change data, such as logging a session, setting a goal or planning tomorrow, usually behind a separate permission. ### Is it safe to connect my health data to an AI assistant? Look for OAuth sign-in rather than pasted passwords or API keys, scoped permissions, short-lived tokens you can revoke, and clear labelling of which assistant wrote what. Only connect assistants you trust. ### Which assistants support MCP servers? Claude, ChatGPT, Claude Code and Cursor, among others, support remote MCP servers, though menus and plan requirements change. Each has its own connect steps. ### Do I need to code to use one? No. With a remote server you paste one URL into your assistant's connector settings, sign in, and approve permissions. --- Canonical: https://revns.com/blog/what-is-a-fitness-mcp-server Index of everything: https://revns.com/llms.txt revns gives training guidance, not medical advice.