wannabe-fit

Whoop-style recovery, strain, and sleep metrics from a cheap watch, computed on your own machine from raw Android Health Connect exports. No cloud account, no subscription, no dashboard, no AI inference. The output is a SQLite database, CSV dumps, Markdown reports, and an MCP server so that Claude Code, OpenClaw, or any other agent can query your data.

Built for a CMF Watch Pro 2 that syncs through the Nothing X app. It works with any watch or app that writes heart rate, sleep, and SpO2 into Health Connect: Samsung Health, Garmin Connect, Zepp, Fitbit, Oura, Polar, Coros, and most budget wearables.

flowchart LR
    W[Watch] -->|Bluetooth| APP[Vendor app]
    APP -->|writes samples| HC[Android Health Connect]
    HC -->|scheduled export zip| GD[Google Drive folder]
    GD -->|rclone copy or manual drop| IN[data/inbox]
    IN -->|vitals ingest| DB[(data/vitals.db)]
    TK[Google Takeout Fit JSON] -->|vitals ingest| DB
    DB -->|vitals compute| M[daily and workout_metrics tables]
    M -->|vitals report| R[data/reports/latest.md]
    M -->|vitals-mcp| C[Claude Code / OpenClaw / any MCP client]
    M -->|vitals sql / export| CSV[CSV]

Why this exists

Watches in the 50 to 100 dollar range collect the same raw signals as a Whoop band: continuous heart rate, sleep stages, SpO2, steps. What they lack is the analysis layer, and their apps keep the raw samples locked in. Health Connect changes that. Since Android 15 the phone can export its whole Health Connect database as a zip on a daily schedule. Inside is one unencrypted SQLite file with every sample every app ever wrote.

This tool reads that file and applies published sports-science formulas to it: Banister TRIMP, Karvonen zones, the impulse-response fitness model, Keytel calorie estimation, sleep debt, and a recovery score built from resting heart rate, sleep, and SpO2. Every number is reproducible from the raw tables, every constant is in one file, and every formula is documented below.

The Google Fit REST API is not an option any more. Google closed it to new sign-ups in May 2024 and ends it at the end of 2026. Health Connect exports are the durable path.

Related MCP server: hybridlog

What you get

A daily table with one row per local day:

A workout_metrics table with one row per session: duration, average and peak HR, percent of HR max, zone minutes, TRIMP, strain, HR-derived kcal next to the device kcal, and 1-minute heart-rate recovery. Plus the raw tables underneath, so you can compute anything else yourself.

Setup

Requirements: Python 3.12+, uv, an Android phone on Android 15 or newer, and a watch whose app writes to Health Connect.

git clone https://github.com/devsrijit/wannabe-fit.git
cd wannabe-fit
uv sync
cp config.example.yaml config.yaml   # then edit: sex, birth year, weight, timezone

On the phone:

1. In the watch app, turn on 24/7 heart rate, SpO2, and sleep tracking. Pick the shortest heart-rate interval it offers. Make sure the app is allowed to write to Health Connect.

2. Settings > Security and privacy > Health Connect > Manage data > Export. Set a daily schedule and pick a Google Drive folder.

Get the zip to your computer. Either:

• Install rclone, run rclone config create gdrive drive scope=drive.readonly, and set rclone.remote_path: "gdrive:<your folder>" in config.yaml. vitals sync then pulls the newest zip before every run. rclone's shared Google client is being retired during 2026, so create your own OAuth client ID as described in the rclone docs.

• Or drop the zip into data/inbox/ by any means. Any file name works.

Then:

uv run vitals doctor   # what the export contains, sample intervals, sources
uv run vitals sync     # pull, ingest, compute, write data/reports/latest.md
uv run vitals today    # the day card

If doctor shows a table the watch should fill but did not, the fix is in the watch app's settings.

Run it in the background (macOS)

scripts/install-launchd.sh          # every 6 hours
scripts/install-launchd.sh 3600     # every hour

This installs a launchd agent named com.wannabe-fit.sync that runs vitals sync and logs to data/sync.log. The sync is idempotent: re-ingesting the same export changes nothing, and rclone only transfers the zip when it changed. On Linux, a cron line running uv run --project <path> vitals sync does the same job.

Web dashboard

web/ is a Next.js site over the same database: recovery ring, hypnogram, zone and load charts, per-workout heart-rate traces, and a glossary that explains every metric with its formula and source. See web/README.md.

cd web && bun install && bun run dev

Interfaces

Command line

MCP server

uv run vitals-mcp starts a stdio MCP server with these tools: get_daily, get_workouts, get_sleep, get_heart_rate, run_sql (read-only), get_schema, get_report, sync.

Register it in Claude Code:

claude mcp add --scope user vitals -- uv run --project /absolute/path/to/wannabe-fit vitals-mcp

Then ask things like "compare my sleep on training days and rest days this month" or "show my heart rate during yesterday's run in 1-minute buckets". Any other MCP client takes the same command.

Files

data/vitals.db is plain SQLite. data/reports/latest.md is always the current report. Both are gitignored, as is config.yaml, so a fork of this repo never carries personal data.

Tables

Raw tables keep epoch milliseconds in UTC. Zone offsets from Health Connect sit next to sessions. daily uses the timezone from config.yaml. Sleep belongs to the day you woke up.

Several apps often write the same thing. Steps, calories, and distance take the largest single-source total per day rather than the sum. Auto-detected sessions from Google Fit are kept unless they overlap a session another app tracked for the same window. Both rules are configurable.

Formulas

The constants live in vitals/metrics.py.

• HR max: hr_max in config, else Tanaka 208 - 0.7 * age. Replace it with a measured peak when you have one. Every zone and strain number depends on it.

• Resting HR (rhr, rhr_method): lowest 15-minute mean of HR inside the main sleep session. Fallbacks in order: lowest 15-minute mean of the whole day, the device's resting HR record, the 5th percentile of the day.

• Baselines: rhr_base30 and rhr_sd30 are the trailing 30-day median and standard deviation, excluding today. Zones and TRIMP use rhr_base30, so one bad night does not move every zone boundary.

• Zones: Karvonen heart-rate-reserve fractions from config.yaml. z1 50-60%, z2 60-70%, z3 70-80%, z4 80-90%, z5 90-100%. Minutes come from each sample's gap to the next, capped at hr_sample_cap_s (default 10 min) in the background and at 60 s inside workouts.

• TRIMP (Banister 1991): sum(minutes * f * 0.64 * e^(1.92 f)) for men, 0.86 * e^(1.67 f) for women, where f is the heart-rate-reserve fraction. Only samples with f >= 0.30 count; sedentary time carries no load. trimp_edwards is sum(zone minutes * zone number).

• Strain (0-21): 21 * (1 - e^(-L / 155)) with L = trimp_edwards. Logarithmic like Whoop, so the last points are the hardest to earn. About 60 minutes in z5 gives 18. This is not Whoop's private formula.

• HR calories (kcal_hr, Keytel 2005): per minute, men (-55.0969 + 0.6309 HR + 0.1988 kg + 0.2017 age) / 4.184, only while f >= 0.30. Validated for steady aerobic work; it overstates strength sessions. The device number sits next to it.

• Fitness, fatigue, form (ctl, atl, tsb; Banister impulse-response): CTL is a 42-day exponential average of TRIMP, ATL a 7-day one, TSB is yesterday's CTL minus ATL. acwr is the 7-day mean over the 28-day mean; above 1.5 is the injury-risk zone in the literature.

• Sleep: sleep_asleep_min is the sum of light, deep, REM, and generic sleeping stages; time in bed is the session length. sleep_need_min = base + 60 * (yesterday's strain / 21) + min(60, 0.5 * 7-day debt). sleep_performance is asleep over need, capped at 100. sleep_consistency loses 25 points for each hour of circular standard deviation in bed and wake times over 7 days (the clock wraps at midnight).

• Recovery (0-100): 0.50 * RHR component + 0.35 * sleep performance + 0.15 * SpO2 component. The RHR component is 70 - 20 * z, where z is today's RHR against the 30-day baseline. The SpO2 component maps the overnight minimum from 88% (0) to 95% (100). If HRV is present the weights become HRV 40, RHR 25, sleep 25, SpO2 10. Bands: green 67+, yellow 34-66, red below 34.

• VO2max (vo2max_uth, Uth 2004): 15.3 * HRmax / RHR. A population estimate, for trend only.

• Workout HR recovery (hrr_1min): mean HR over the last 15 s of the session minus mean HR 50-75 s after it. Needs the watch to keep sampling after the workout ends. Above 20 bpm is good, below 12 poor.

Adapting it

• Another watch: nothing to change if the app writes to Health Connect. Run vitals doctor on the export to see which tables are filled. If the watch provides HRV, recovery picks it up on its own.

• Older history: vitals ingest also reads a Google Takeout export of Google Fit (the zip or the Fit/All Data folder). Fit auto-detected walks arrive as workouts with source Fit.

• Different formulas: every constant is in vitals/metrics.py, and compute rebuilds the derived tables from raw samples, so changes apply to all history at once.

• Testing without a phone: uv run python scripts/make_fixture.py /tmp/fixture.zip builds a synthetic 35-day export in the real Health Connect layout. Ingest it from /tmp, never from data/inbox, or sync mixes synthetic rows into your real database.

Limits

• Health Connect has no data type for vendor stress scores, so they do not export.

• Background HR interval is whatever the watch app allows. doctor prints the real median gap.

• Health Connect keeps 30 days of most data on the phone by default. Daily exports snapshot everything, so nothing is lost while the schedule runs.

• The recovery score without HRV leans on resting HR, sleep, and SpO2. It is a trend instrument, not a diagnosis.

License

MIT.