How can you track calories with ChatGPT or Claude?

Because a chat has no total. You can tell ChatGPT every day what you ate and it will answer kindly with estimates, but it measures no portion, checks nothing against a nutrition database, counts toward no target and forgets the history once the context fills up. Ask after three weeks "what was my average protein?" and you get a recollection of whatever is still in the chat, not an analysis.

Conversely, a tracking app is a poor conversation partner: it shows numbers but does not explain them, does not plan your evening and answers no question nobody built a button for. The combination fixes both, provided the assistant may read the real numbers.

Kairo exposes an MCP endpoint; MCP is the open standard through which AI assistants attach tools. In Claude (claude.ai or Desktop) and ChatGPT you add Kairo as a connector with the endpoint address; for Claude Code, Codex CLI and Gemini CLI there is a one-line command each. On first access the browser opens: you sign in with the same Apple or Google account as in the app and confirm that this assistant may access your data.

If you have no Kairo account yet, the page points you to the App Store. The connection is for existing users, not a standalone service. Step-by-step instructions per assistant and copy-ready commands are at mcp.kairocalories.com.

Ten things, eight of them read-only. Overview (profile, goals, streak, current status), daily macros for the last N days with averages, a meal list with name search, details of a single meal including micronutrients, weight history with target and change, micronutrients per day, food statistics (top foods, weekday versus weekend) and your weekly body score. Answers are compact and rounded so your assistant's context does not fill up with raw dumps.

Two tools write: log a meal from text (it goes through the same analysis in Kairo as an in-app text scan, including the reference cross-check) and log your current weight. Both are described so that the assistant uses them only on your explicit instruction. A duplicate guard stops an assistant that retries from creating the same meal twice.

The ones no app has a button for. "Weekly review: what went well, where did I miss?" "Plan the rest of my day with what is in the fridge, 40 g of protein still open." "How does my weekend differ from my weekdays?" "Which five foods give me the most sugar?" Two of these (the weekly review and the rest-of-day plan) ship as ready-made prompts that appear in your assistant as entry points.

What is not worth it: sending photos to the assistant. The photo scan with portion estimation and label lookup lives in the app; the assistant gets the finished numbers. If you want to know how good those are, the benchmark is on the accuracy page.

The connection runs on your own login token, issued after the browser sign-in. Kairo's server uses no master key for it: every request hits the same per-user access rules as the app itself, so only your own data. Each tool call is logged (which tool, whether it succeeded, how long it took), never the contents of your request. A limit of 30 calls per minute keeps an assistant stuck in a loop from doing damage.

One honest limit: the token is a full user token. A compromised assistant could do with it whatever the app may: strictly confined to your account, but not read-only. So connect only assistants you trust, and revoke the connection under "Connections" when you no longer need it. What Kairo does with data in general is covered in Is there a Cal AI alternative built in Europe with GDPR hosting?.