Garmin Local MCP

I wanted to reason over my own health data, not glance at a dashboard. That means questions that span months, like "how does my sleep correlate with training load?", and it means the data has to survive Garmin breaking its own login, which it did in March 2026 and took most of the Garmin tooling down for weeks.

So garmin-local-mcp keeps your history in a local warehouse you own, does the analysis there, and hands your AI assistant small answers instead of raw data.

What it does

• Sync once, analyze forever. Incremental sync into a directory you own: immutable raw JSON snapshots plus a SQLite database.
• Analysis on your machine, compact answers. Trends, correlations, personal baselines and anomaly detection are computed locally and returned as small tables, typically under 2 KB, so nothing floods the model's context.
• Keeps working when Garmin breaks. An API outage pauses new syncs only. Every question over history you already have keeps working.
• Ingest with no login at all. A decoder for Garmin's undocumented wellness FIT files (sleep score, HRV, skin temperature, sleep stages, naps) reads a manual export with no credentials. I haven't found another Garmin MCP that decodes the wellness export.
• Correlation with lag. "Training load raises my resting heart rate the next day" is a question with an answer: it scans lags from -7 to +7 days and tells you when the strongest one might be chance.
• Curated, not sprawling. 13 composable tools, not 110.

New in October 2026: when should I wake up?

It started with a plain question to Claude: based on my Garmin, what's my circadian rhythm, and what's the best time for me to wake up? The first answer was a reasonable formula. The second question was the better one: is that formula actually right for me?

The new circadian tool answers both, as three separate estimates:

• Your chronotype. The standard sleep-science measure (from the Munich Chronotype Questionnaire): the middle of your sleep on free days, corrected for the catch-up sleep you take after short workdays.
• Your sleep need, taken from how long you sleep when no alarm is set.
• A wake window that fits both. In plain terms, it's roughly when you wake up on free days once you take catch-up sleep out of the picture.

Then it checks its own work, which is the part I care about most:

• How sure is it? Every estimate comes with a range, so you can tell whether "07:05" is precise or just the middle of a wide spread.
• Does the answer hold up? It recomputes everything over the last 90, 180 and 365 days and a few other reasonable settings. If your schedule changed recently, it says so.
• Did it actually work? It trains on months of your history and tests on the next month it hasn't seen, again and again. The question is whether waking near the suggested time went with better Garmin scores than sleep length alone would predict, and whether that's more than just keeping a regular schedule. It reports one of four levels, from consistent to unsupported, and it is happy to tell you the formula didn't hold.

Chronotype on its own isn't new; other tools report it. What I haven't found anywhere else is advice that comes with its uncertainty and gets tested against your own data before you're asked to trust it. It's also careful about what it can't say: Garmin's scores are Garmin's own formulas, not a lab measurement, so the strongest honest claim is "associated with higher Garmin scores," never "better for your body."

The design went through four rounds of outside review before I wrote a line of code. The best change from those reviews was the regularity check: a wake time can look good just because the nights near it were regular ones, so the tool tests for that separately.

What you can ask

Once your history is synced, Claude can answer questions like these in a call or two:

• "Am I a morning person, and what time should I wake up?"
• "How does my sleep score correlate with next-day resting HR?"
• "What were my anomalous HRV days this quarter?"
• "Show weekly training load vs sleep for the last three months."

Try it without a Garmin

garmin-local-mcp demo builds a synthetic 180-day store with real structure planted in it: an illness week, a next-day training effect, and a body clock set to 03:45. The tools find all three. No account, no login, no network.

How it works

It's a Python package (pip install garmin-local-mcp). You log in once (MFA supported; tokens persist locally), backfill your history with a resumable, rate-limit-friendly sync, and register the server with any MCP client. With Claude Code:

claude mcp add --scope user garmin -- garmin-local-mcp serve

Claude Desktop users can install it in one click from the .mcpb file on the GitHub releases page. Everything lives in one directory you own (~/.garmin-mcp by default). Only two of the thirteen tools write anything, and only inside that directory. MIT licensed.