FitSnap Coach

FitSnap Coach is a bilingual, local-first AI fitness coach MVP that turns goals, meals, workouts, movement media, and recovery metrics into a daily action plan. It runs as a static website, stores user data in the browser, and includes a browser-side agent workspace for nutrition, training, form, and recovery tasks.

Open the live demo or run it locally in one command.

Why Star This Repo

• Complete AI fitness product loop: goal setup, calorie and macro targets, meal logging, training plans, form checks, recovery scoring, trend charts, and an agent action queue.
• Local-first privacy model: IndexedDB is the primary database, with LocalStorage fallback. User media and health data are not uploaded to a server.
• Real pose pipeline path: TensorFlow.js and MoveNet SinglePose Lightning can run in the browser for keypoint detection, with a rule-based fallback when the model or network is unavailable.
• Deploys like a static app: no backend, no build step, no API key required for the MVP.
• Useful reference architecture: a compact vanilla JavaScript example for AI product prototyping, health-tech UX, and browser-only persistence.

Run Locally

Open index.html directly in a browser. If the browser restricts IndexedDB on file://, run this from the project directory:

python3 -m http.server 4173

Then open:

http://127.0.0.1:4173/index.html

Feature Map

Area	What is included
Goals	Body metrics, target weight, activity level, training experience, equipment, injuries
Nutrition	Calorie and macro targets, photo or text meal logging, editable estimates
Training	7-day plan generation, workout completion tracking, recovery-aware advice
Form analysis	Photo/video upload, MoveNet keypoints when available, rule-based scoring fallback
Live motion	Camera preview, skeleton overlay, refreshed form feedback
Recovery	Simulated Apple Health authorization, JSON/CSV import, sleep, HRV, RHR, SpO2, steps, load
Agent	Local observe/reason/act loop, task queue, section deep links, persisted messages
Trends	Weekly and monthly charts for calories, protein, workouts, readiness, form, uploads

Pose Model Integration

The form analysis pipeline uses a hybrid approach:

uploaded photo/video
-> TensorFlow.js + MoveNet SinglePose Lightning
-> body keypoints
-> angle, symmetry, torso-lean, knee/ankle tracking signals
-> rule-based form scoring
-> coach-readable feedback

The model is loaded dynamically from jsDelivr only when the user clicks Load pose model or runs a form analysis. If TensorFlow.js, the pose model, or reliable keypoints are unavailable, the app falls back to local rule-based analysis so the product remains usable offline.

Live camera mode uses navigator.mediaDevices.getUserMedia and MoveNet in the browser. It renders a skeleton overlay and live feedback, but it does not save every frame to history. To persist a form-analysis record, upload a photo/video and run Generate form feedback.

AI Coach Agent

The agent workspace uses a local rule-based loop:

observe local profile, meals, training, form, health, and trend data
-> reason about the highest-impact constraint
-> generate nutrition, training, recovery, or form tasks
-> let the user open the relevant section or mark the task done
-> persist the messages and tasks in IndexedDB

This gives the product an agent-like workflow without requiring a backend or API key. A production build can replace the local reasoning layer with an LLM call while keeping the same context builder, task schema, and safety boundaries.

Local Data Model

The IndexedDB database is named fitsnap-coach-db.

Object stores:

• meta
• profile
• nutritionTargets
• meals
• workoutPlan
• workoutCompletions
• formAnalyses
• healthSnapshots
• mediaAssets
• agentTasks
• agentMessages

Deployment

Production URL:

https://fitsnap-coach.vercel.app

Vercel

1. Push this folder to a GitHub repository.
2. Import the repository in Vercel.
3. Use the default static-site settings. No build command is required.
4. Vercel will serve index.html over HTTPS.

The included vercel.json sets browser security headers and allows camera access from the same origin for live form checks.

Netlify

1. Push this folder to a GitHub repository or drag the folder into Netlify Deploys.
2. Netlify should publish the project root.
3. No build command is required.

The included netlify.toml publishes the static root and sets matching security headers.

Roadmap

• Add a real social preview image to the GitHub repository settings.
• Add a license before asking external contributors to reuse or fork the code.
• Add hosted LLM calls for the agent while preserving local privacy boundaries.
• Add optional auth, cloud sync, and media storage for cross-device continuity.
• Add automated smoke tests for local persistence, language switching, and core agent flows.

Growth Notes

See docs/star-growth-playbook.md for GitHub metadata, launch checklist, and ready-to-post English and Chinese copy.

Health Boundary

FitSnap Coach does not provide medical diagnosis and does not directly measure cortisol. The recovery score uses proxy signals from health and training data to estimate stress load trends. If heart rate, blood oxygen, sleep, or physical symptoms remain abnormal, users should consult a qualified professional.