peak-lab
A Claude Code plugin that turns your Garmin Connect data into a multi-agent
training analysis. It pulls the past couple of months of activities,
wellness, and training-status data, dispatches three specialist subagents in
parallel (metrics, activity, physiology), synthesizes their outputs, and
renders a self-contained HTML dashboard.
peak-lab is a research preview and a Claude-Code-native port of
leonzzz435/garmin-ai-coach
(MIT) — the same three-expert architecture, re-expressed with Claude Code
skills, subagents, and two small CLIs (bin/fetch-garmin, bin/garmin-call)
as the data layer.
Not Garmin Coach. Garmin Coach is Garmin's own adaptive training
feature. peak-lab is an unaffiliated third-party tool that reads data from
your Garmin Connect account.
Quickstart
You need Claude Code, a Garmin Connect account, and uv (the
Python runner peak-lab's CLIs use).
- Install
uv:
curl -LsSf https://astral.sh/uv/install.sh | sh
- Clone the plugin:
git clone https://github.com/shlomihod/peak-lab.git
- Make a training-project folder and launch Claude Code from it:
mkdir ~/my-training-project && cd ~/my-training-project
claude --plugin-dir /path/to/peak-lab
- Bootstrap: run
/peak-lab:setup. It prints the path to the bundled
bin/garmin-auth script — run that once in a terminal. It prompts for
your Garmin email, password, and an MFA code (if enabled), then saves
OAuth tokens to ~/.garminconnect/garmin_tokens.json. Your password is
never stored. (Re-run garmin-auth --verify to check tokens, --force
to log in again.)
- Analyze: run
/peak-lab:analyze. The first run takes a couple of
minutes while uv resolves dependencies; later runs are ~15–30 seconds.
--plugin-dir loads peak-lab for one session — the intended path while it's
unpublished. Once it's in a plugin marketplace, installation becomes
/plugin marketplace add <marketplace> then /plugin install peak-lab@<marketplace>.
The commands
/peak-lab:setup — one-time bootstrap. Checks prerequisites, walks you
through Garmin authentication, and captures a short athlete profile (goals,
constraints, races) in profile.md. Re-run any time to update the profile./peak-lab:analyze — pulls ~60 days of Garmin data (activities, sleep,
HRV, training status/readiness, body battery, plans, PRs, race predictions)
into raw.json, dispatches the metrics, activity, and physiology
expert subagents in parallel, synthesizes their findings, and writes
analysis.html + analysis.json. Open the HTML in a browser.
Where your data lives
peak-lab writes every file into the current working directory — the
folder you launched Claude Code from. That folder is your training project:
profile, raw data, analyses, and plans all live there. No env vars, no
fallback paths.
Keep your training project separate from the plugin source — don't run
peak-lab from inside the cloned peak-lab/ folder. A typical project after an
analyze run:
my-training-project/
├── profile.md your athlete profile (goals, constraints, races)
├── garmin-call.log audit log of every Garmin API call
└── runs/
├── 20260413T142123Z/ one analyze run
│ ├── raw.json full Garmin pull (~20 MB)
│ ├── analysis.json
│ └── analysis.html open this in a browser
└── latest → … symlink to the most recent run
If you git-track the project, ignore the bulky raw data in .gitignore:
runs/*/raw.json
garmin-call.log
Garmin OAuth tokens live separately at ~/.garminconnect/garmin_tokens.json
— global, written only by garmin-auth, reused across all your projects.
Configuration
peak-lab uses Claude Code's userConfig mechanism — you're prompted when the
plugin is enabled:
| Key | Default | Meaning |
|---|
analysis_window_days | 60 | How many days of data /analyze pulls. |
How it works
