hybridlog

hybridlog

Self-hosted MCP server on your own Garmin data. It mirrors everything Garmin answers into a database you control, computes the models the watch does not (training load, muscle freshness, readiness in context), and hands all of it to a language model over stdio or HTTP, so the things Garmin cannot measure can be dictated into the chat instead of typed into a form. A dashboard comes with it: the body map, the training load, and the login the connector authenticates against. There is no hosted instance, and that is the point: it is a health record, and the model reads it through a role that can only read.

Running, lifting or both on the same day: the models do not care which. Sessions that alternate running with station work get a lap-by-lap breakdown of their own, which is the shape of a HYROX race.

There is a demo to walk through before installing anything: sign in as demo@example.com with the password demo-demo-demo. It runs on generated data and puts itself back every night.

Every number in these three is generated by fetcher/seed_demo.py. The interface is English and ships a German translation; it follows the browser language unless the profile says otherwise.

Read this before you rely on it. The fetcher talks to Garmin

Connect's *unofficial* web API. Garmin can change or close it without

notice. This project is not affiliated with or endorsed by Garmin.

Nothing it computes is a medical statement: readiness, HRV bands, load

ratios and the coach texts are training aids, not diagnoses. It is a

personal project published in the hope it is useful, with no promise of

support or a stable interface.

What it is

• MCP server: thirteen tools and one prompt over laravel/mcp, in two transports. Local (stdio) for Claude Code and Claude Desktop, hosted (streamable HTTP with OAuth 2.1, PKCE and dynamic client registration) for claude.ai, ChatGPT, Mistral's Le Chat (renamed Vibe), LM Studio and anything else that speaks it.
• Dashboard: Laravel + Alpine + Chart.js. It draws the body map and the training load, and it is also the login the connector authenticates against and the place where you set what the chat may see.
• Fetcher: Python (fetcher/), pulls daily metrics and activities via python-garminconnect into one athlete's schema of a PostgreSQL database, which both of the above read through a connection that may only read it, and only that athlete's. Each answer is also kept whole beside the columns read out of it, so a field nobody has built a column for is still there to be asked about, and a day Garmin Connect no longer serves is still in the mirror.

What *no hosted instance* means in practice: the Garmin session sits in a schema of your own database that no reader role can reach, so the SQL a model writes cannot ask for it, and no operator but you holds it. Past Garmin itself, two optional things reach out at all: the weather, which takes a pair of coordinates you fill in yourself, and notifications, where the push service is woken by an empty POST and never carries what it was about.

The tools

Thirteen tools and one prompt. The right-hand column is the switch at /connect that gates each one; everything is on by default.

Most of them answer a question the app was built around. Two do not: describe-schema hands the model the mirror's own tables with the comments that say what each column means, and query-health-data runs the SELECT it writes from that. It is the difference between a set of reports and a database somebody can think in, and it is what makes keeping the raw answers worth the disk: a field nobody has promoted to a column is still one question away.

Reading is the whole of it, with one documented exception. Symptoms are the only thing the chat may write, and they go to the app's own schema, never into the Garmin mirror. Free-form SQL runs through app/Garmin/ReadOnlyGarminQuery: a single SELECT or WITH, a keyword blocklist, a read-only transaction, a row cap, on a connection switched into a role that holds SELECT on one athlete's schema and nothing else.

Quickstart (Docker, demo data, no Garmin account)

git clone https://github.com/Imgrund/hybridlog.git

cd hybridlog cp .env.example .env docker compose up -d

That brings up PostgreSQL, the dashboard, a queue worker and a scheduler. The published image covers amd64 and arm64, so the first start downloads rather than compiles. After a git pull, plain up -d keeps running the image it already has: the new code arrives only with --build.

Then fill the mirror with 120 days of plausible data, create the account you log in with, and open the dashboard:

docker compose exec app /opt/fetcher/bin/python fetcher/seed_demo.py

docker compose exec app php artisan app:create-user you@example.com --admin open http://localhost:8080

There is no sign-up page, on purpose: a login nobody can register at has no surface to attack. For anybody but yourself, hand over a link instead of a password:

docker compose exec app php artisan app:invite them@example.com --name="Them"

It prints a one-time link, good for seven days (--days), on which they set their own password. --admin marks the installation owner, who is the account the local stdio transport acts for. Every account keeps its own profile, permissions, symptom log, Garmin sign-in, notifications and mirror, and sees none of anybody else's.

DEMO_MODE=true turns an installation into a shop window instead: one shared account, everything that would reach out of it closed (the Garmin sign-in above all), and php artisan demo:reset putting it back nightly.

A shop window is also the one place worth counting visitors in, and UMAMI_SCRIPT_URL plus UMAMI_WEBSITE_ID render the script tag of an Umami instance you host yourself: page views without a cookie and without an identifier that follows anyone off the site. Both lines are needed, and with either empty, which is the default, nothing is rendered at all and no page requests a host but this one. It is not tied to DEMO_MODE, so a private installation stays free of outside scripts by simply leaving them alone.

Connect an AI

The server reads the same mirror the dashboard draws from, so a chat and the page never disagree. What the page cannot do is answer a question it was not built for: how this week stands against the one before, whether today is a rest day, what a niggle means for tomorrow.

[!IMPORTANT]

This hands a language model your health record. It reads through a role

that may only read, and only your own schema, but it reads all of it.

Which parts is yours to set at /connect, per switch, and the switches

take effect on the next tool call rather than the next reconnect.

One address, the same for every client:

https://<your-domain>/mcp/garmin

Claude, ChatGPT, Langdock, Le Chat and LM Studio all take it and run OAuth against the dashboard's own login, so no client ever sees a password and /connect can cut any of them off again. Claude Code and Claude Desktop can skip the deployment entirely and talk to the repository over stdio. docs/connect-ai.md has the steps per client.

Whichever way it connects, it gets the same thirteen tools and the same switches.

Things worth asking, once it is connected:

How did my training week compare with the one before?

Am I ready for a hard session, or do I need a rest day?

Which muscles are fresh, and what should I train today?

Where did my pace go in Saturday's race, and how much of the clock was

station work?

My left knee hurt on the box jumps.

Setup with your own Garmin account

Sign in under *Garmin* in the account menu, or at /connect/garmin: email, password, and the MFA code if Garmin asks for one. What is stored is an OAuth token pair, never the password. The sign-in runs on a queue worker, so one has to be up.

A first sign-in fills the mirror by itself: a ninety-day backfill on the queue, roughly a quarter of an hour, with the page filling in as the history lands. Ninety days is about the minimum for the models to say anything, since the HRV baseline needs three weeks of nights and the load ratios a rolling six weeks. From then on the scheduler fetches three times a day and the *Fetch from Garmin* button fetches on demand.

Every command works on one athlete (--tenant , the owner where left out). Details, backfills and the manual login are in docs/install.md; how to record so the data is worth reading is in docs/recording.md.

Documentation

• docs/install.md: installing without Docker, deploying to a platform, backup.
• docs/connect-ai.md: every client's steps, OAuth and discovery, running the MCP server on a public host.
• docs/recording.md: recording rules, weather, notifications.

Architecture notes

-