A private, self-hosted health dashboard for one person.
Sleep, recovery, movement, food, hydration and longevity in one calm iPhone PWA,
fed by your Fitbit / Google Health data, with an AI coach that can log for you,
and sip reminders that cover every screen through a small desktop app.
Commercial health apps keep your data on their servers and hide how their scores work. Demel runs on your own machine (a mini PC, a home server, a VPS behind Tailscale), stores everything in one SQLite file, and shows exactly how every number is calculated. There is one user, one password and no tracking.
All screenshots use the built-in demo mode with synthetic data. Nothing shown belongs to a real person.
Today
- Three score rings: Strain, Recovery and Sleep. Tap a ring for its detail page with a 7-day trend, its inputs and the method behind it.
- A daily physiology engine built on per-minute heart rate: heart-rate zones, Edwards TRIMP strain with a target strain based on your 14-day load and today's recovery, HRV/RHR-based recovery against your 30-day baseline, plus Stress, Energy, heart-rate recovery and training load.
- Short AI insight cards written from today's numbers (optional, only with an OpenRouter key).
- Vitals at a glance: HRV, resting heart rate, sleep and SpO₂, with a 14-day chart.
Journal, Fitness and Biology
- Journal: one-tap daily habits (yes/no), caffeine and alcohol amounts, illness, shown next to the following night's sleep and recovery.
- Fitness: training load, cardio fitness and per-session detail with zones and energy.
- Biology: a weekly biological age from the last 4 weeks across five categories (Sleep, Activity, Fitness, Lifestyle, Blood). Each factor comes from published hazard ratios, converted to years with the Gompertz mortality law, with an overlap correction and a confidence score per factor.
Health
- Overview of all wearable measurements: HRV, resting HR, SpO₂, breathing rate, skin temperature, VO₂ max, weight, steps and blood pressure.
- Sleep with stages and efficiency, 90-day trends and journal patterns.
- All data: a catalogue of 42 Google Health data types with daily summaries, every numeric field, per-type history and the original records (stored losslessly and compressed).
- Age & patterns: a VO₂ max fitness age (FRIEND 2022 reference), sleep debt, circular bedtime regularity, moderate-intensity minutes, protein per kg and energy balance, plus correlations between sleep, activity and HRV.
- Optional Clinical PhenoAge from 9 blood markers (Levine 2018), and storage for epigenetic clock results (Horvath, GrimAge, DunedinPACE and others).
- Daily check-in for mood, energy, stress, caffeine, alcohol, meditation and illness.
Food & hydration
- Food diary with calories, macros and fiber, a food library and saved portions.
- Describe a meal or snap a photo and the AI estimates the portion; you always review it before saving.
- Hydration with one-tap Sip / Half glass / Glass buttons. Drops fall into a human-figure vessel that fills up.
- Sip reminders at random times across your reminder hours: the time left divided by the sips left, times a random factor between 0.5 and 1.5, never sooner than 5 minutes. A glass or a bottle you log reshapes the rest of the day.
- A full-screen reminder (Done / In 5 min / Skip) covers the open app, the phone notification opens it, and the desktop app covers every monitor. Answering on one device clears it everywhere, and the phone stays quiet while you are at a computer.
Training
- Workouts with exercises and sets, RPE, a stopwatch and a rest timer.
- Reusable plans, an exercise library and a fully paginated history.
- Strength volume chart; Fitbit activities sync in automatically.
AI coach
- A chat that sees a compact 30-day summary of your data and answers in the language you write in.
- It can log for you: "I just had two eggs and a coffee" becomes journal entries through validated tool calls (water, meals, workouts, check-ins, measurements; only today and the last 7 days).
- Every reply shows its cost in USD in the corner of the bubble, plus the total for the conversation.
- Works with any OpenRouter model. The recommended default is
anthropic/claude-haiku-5.5: cheap, fast, reliable at tool calls and able to read food photos.
Everything else
- Google Health OAuth with token refresh, pagination, manual sync of up to 90 days and an automatic hourly sync. The last successful sync time is always visible.
- Web Push notifications with your own VAPID keys: a morning summary, an evening check-in, hydration and movement reminders.
- Installable PWA with an offline read-only copy (opt-in) and automatic updates when a new build is deployed.
- JSON export and import, plus daily consistent SQLite backups kept for 14 days.
iPhone PWA / browser desktop app (tray, every monitor)
│ │ SSE /api/hydration/events, device token
│ HTTPS (e.g. Tailscale Serve)
▼ ▼
FastAPI + Uvicorn ── serves the built frontend and the /api
│ │
│ ├── scheduler: hourly Google Health sync, push reminders, daily backup
│ ├── Google Health API (read-only OAuth)
│ ├── OpenRouter API (only with your key and explicit consent)
│ └── Web Push (Apple / Google / Mozilla push services)
▼
SQLite (WAL) ── entries, daily Google archives, settings, encrypted secrets
- Frontend: vanilla TypeScript built with Vite. No framework; views are template strings with strict escaping, behind a strict Content-Security-Policy.
- Backend: FastAPI with pydantic validation on every write. One process; no Redis, queue or external database.
- Scores: plain Python in
backend/engine.py(daily physiology),backend/bioage.py(biological age),backend/scores.py,backend/analytics.pyandbackend/age.py, covered by tests. Every score's method and references are also shown in the app. Formulas and sources are in docs/SCIENCE.md. - Design: a dark “bevel” design system with accents taken from the logo. Details in docs/DESIGN.md.
Requirements: Python 3.12 and Node 22.
git clone https://github.com/greeenos/demel.git && cd demel
python3.12 -m venv .venv
.venv/bin/pip install -r requirements.txt
npm ci
.venv/bin/python bootstrap.py # creates .env and ACCESS.txt (your first password)
npm run build
.venv/bin/python run.py # http://127.0.0.1:8198bootstrap.py runs once: it generates a random password (written to ACCESS.txt, mode 0600), its PBKDF2 hash and a Fernet encryption key in .env. It never overwrites an existing configuration. Use VITAL_URL=https://your-host when creating it to set the public origin. Change the password in Settings → Security after your first login.
Just want to look around? Open the app and tap Explore the demo. The demo runs entirely in the browser with synthetic data.
.venv/bin/python bootstrap.py # or copy an existing .env
docker compose up -d --build
docker compose logs --tail 50 vitalThe container:
- runs as UID 10001 with no Linux capabilities,
no-new-privilegesand a read-only root filesystem, - has 256 MB RAM and 1 CPU, a health check on
/api/health, and is bound to127.0.0.1:8198, - writes only to
./data.
Put HTTPS in front of it. With Tailscale this is one command:
tailscale serve --bg --https=10456 http://127.0.0.1:8198Set APP_URL in .env to the exact public origin (scheme, host and port). It is used for the CSRF origin check, the OAuth redirect URI and the secure cookie flag.
- In Google Cloud, create a project, enable the Google Health API and create an OAuth client of type Web application.
- Add the redirect URI shown in Settings → Fitbit / Google Health (
<APP_URL>/api/google/callback) and add your Google account as a test user. - Paste the Client ID and secret into Settings, click Connect Google account and then Load 90 days.
Demel only requests read-only scopes: activity and fitness, sleep, health metrics, nutrition, irregular rhythm, ECG, profile and settings. Partial failures keep the successful parts, and every data type shows whether it is available, empty, missing a permission or unsupported by your device.
Add an OpenRouter API key and a model in Settings → OpenRouter AI, then turn on consent. Without both, no AI request is ever sent. Each question sends a 30-day summary, the last week of meals, workouts and check-ins, and recent chat turns to OpenRouter and the chosen provider; food photos are sent only when you use recognition. Responses request usage.include, so the exact cost of each message is stored next to it.
Open the app in Safari → Share → Add to Home Screen, open it from the icon, then go to Settings → Smart notifications → Enable on this device and send a test. iOS Web Push needs the app installed on the Home Screen. Delivery times are best effort.
Set the reminder hours, days and sip size in Settings → Hydration schedule and turn on Hydration reminders under Smart notifications. When a sip is due, a full-screen reminder covers the open app, and tapping the phone notification opens it (/?overlay=1; Enter = done, Esc = in 5 min). Answering on any device clears the reminder everywhere.
The desktop app lives in the tray and covers every monitor until you answer. While it sees you using the computer, Demel skips the phone push. It is a ~2–5 MB installer built with Tauri; the source is in desktop/.
| Platform | Download from Releases |
|---|---|
| Windows | Demel_x.y.z_x64-setup.exe, or the portable .exe |
| macOS (Intel + Apple Silicon) | Demel_x.y.z_universal.dmg, then run xattr -cr /Applications/Demel.app once (unsigned build) |
| Linux | Demel_x.y.z_amd64.AppImage or .deb (the tray icon needs libayatana-appindicator3) |
On first launch, enter your Demel address and password. The password is exchanged once for a device token and never stored. Tray menu: I took a sip, +250 ml glass, +500 ml bottle, Remind everywhere now, Open Demel, Start at login.
- Single user. Login uses PBKDF2-SHA256 (600k iterations), an HttpOnly SameSite cookie stored as a SHA-256 hash server side, and rate limiting.
- The desktop app gets a device token (a 365-day session, sent as
Authorization: Bearer). Its windows hand the token over in the URL fragment, which never reaches the server, and swap it for the usual cookie. Changing the password revokes every session and device. - Every non-GET request is checked against
APP_URL(CSRF), and request bodies are capped at 9 MB. - Integration secrets (OpenRouter key, Google OAuth client and tokens, VAPID key) are encrypted at rest with Fernet.
- Strict CSP (
script-src 'self'),X-Frame-Options: DENY,no-referrerandnosniff. - No analytics, no third-party scripts, no fonts from a CDN.
Keep .env safe and backed up separately: without its ENCRYPTION_KEY, the stored integration secrets cannot be decrypted.
data/vital.sqlite uses WAL. A consistent copy made with the SQLite backup API is written to data/backups/ every day and kept for 14 days. A backup on the same disk does not protect against disk failure, so copy it elsewhere.
To restore:
- Stop only the app.
- Back up the current
data/. - Put the chosen backup in place as
data/vital.sqlite. - Remove the old
-waland-shmfiles and restore ownership to UID 10001. - Start the app again and check
/api/health.
The JSON export (Settings → Your data) restores health entries by ID; profile, chat and tokens come back only from a full database backup.
.venv/bin/pip install pytest playwright && .venv/bin/playwright install chromium
.venv/bin/pytest -q # backend: scores, PhenoAge, auth, sync, AI tools, regressions
npm run build # type check + production buildThe browser checks in scripts/ need an isolated test backend. They create test records, so never point them at your real database:
DATA_DIR=data-browser APP_URL=http://127.0.0.1:8208 PORT=8208 \
APP_PASSWORD_HASH=... ENCRYPTION_KEY=... .venv/bin/python run.py
VITAL_TEST_URL=http://127.0.0.1:8208 VITAL_TEST_PASSWORD=... .venv/bin/python scripts/browser_check.py
VITAL_TEST_PASSWORD=... .venv/bin/python scripts/sip_browser_check.py # full-screen reminder, device token, SSESet PW_CHROMIUM to any installed Chromium (e.g. Brave) if Playwright's own browser is not installed. scripts/screenshots.py regenerates the README screenshots from the demo. The desktop app is built on GitHub Actions; see desktop/README.md.
backend/
main.py FastAPI app: auth, API routes, AI chat and tool loop
models.py pydantic models for every entry type
scores.py sleep / recovery / strain and daily aggregation
analytics.py fitness age, sleep debt, regularity, correlations
age.py Clinical PhenoAge and the study catalogue
google_health.py OAuth and sync
google_data.py lossless daily archive for 42 data types
coach_tools.py tools the AI coach may call to log entries
hydration.py sip-reminder schedule, desktop presence
notifications.py Web Push and the background scheduler
db.py SQLite helpers, encrypted secrets, backups
frontend/
main.ts state, actions, forms, motion system
views.ts screens
insights.ts Google data and long-term health views
sip.ts full-screen sip reminder (/?overlay=1)
water.ts hydration vessel and drop physics
ui.ts components, charts, wordmark
icons.ts generated Streamline duo icons
style.css the design system
public/ logo, PWA manifest, service worker
desktop/ tray app for Windows, macOS and Linux (Tauri, GPL-3.0)
docs/ design system and scientific method
tests/ pytest suite
scripts/ browser checks, screenshots, icon export
.github/workflows/ desktop builds and releases
- Icons: Streamline Flex (CC BY 4.0), recoloured.
- Wordmark: Unbounded (SIL Open Font License 1.1), converted to outlines.
- Fitness reference: Kaminsky et al., FRIEND registry, Mayo Clinic Proceedings 2022.
- PhenoAge: Levine et al., Aging 2018.
- Desktop app: started as Drinky, itself a fork of BreakTimer by Tom Watson (GPL-3.0).
Demel is a personal wellness tool, not a medical device. Its scores are transparent heuristics, not diagnoses. Do not change medication, diet or training because of a number in this app without talking to a qualified professional. Lab results must be interpreted by a doctor.
No open-source license has been chosen yet, so the code is © the author, all rights reserved. Third-party assets keep their own licenses (see Credits). The desktop app in desktop/ is GPL-3.0-or-later (see desktop/LICENSE.md).













