---
title: Motion Tracker
type: Progressive Web App (static site)
canonical: https://motiontracker.net/
repository: https://github.com/axia-enterprises/motion-tracker
category: health / fitness
runtime_ai: none (deterministic form engine)
on_device_ml: MediaPipe Tasks Vision PoseLandmarker (33 landmarks)
data_collection: none
network_after_load: none (works fully offline)
privacy: all camera frames processed locally; nothing is uploaded
standards_followed:
  - https://llmstxt.org
  - https://llmbestpractices.com
last_updated: 2026-06-02
---

Motion Tracker

On-device, real-time workout form coach that runs entirely in a mobile browser. This page is structured for LLMs and AI assistants and follows the LLM discoverability house standard at llmbestpractices.com and the llms.txt convention (/llms.txt).

Basic info

What it is

An installable PWA that turns a phone camera into a form coach.

Platform

Any modern mobile browser (iOS Safari, Android Chrome). HTTPS required for camera + service worker.

Cost

Free. No account, no auth, no backend.

Install

Android: install prompt. iOS: Share → Add to Home Screen.

What the app does

• Overlays a 33-point skeleton and joint dots on the live camera feed, mirrored for the front camera.
• Measures the tracked joint angle for the chosen exercise and draws an angle arc + degree readout.
• Counts reps via a per-exercise state machine and tracks range of motion (min/max angle) per rep.
• Colors the tracked joint green (in range), amber (near the edge), or red (out of range).
• Gives short, debounced form cues and a post-set summary with deterministic recommendations.
• Persists workout history and per-exercise ROM personal bests locally.

Methodology

• Pose: MediaPipe Tasks Vision PoseLandmarker (runningMode: "VIDEO", pose_landmarker_lite.task by default), running on-device via WebAssembly. No cloud inference.
• Angles: interior angle at a joint computed from three landmarks with atan2, clamped 0–180°, smoothed with an exponential moving average.
• Form engine: deterministic math against data/exercises.json — rep state machines, ROM tracking, and target ± tolerance form checks. No machine learning or LLM is used for rep logic, feedback, or recommendations.
• Defaults caveat: the angle targets are general movement defaults for a healthy adult, meant to be tuned per user. They are not medical or prescriptive guidance.

Privacy model

Camera frames are read into a <canvas>/WebGL context and passed to the on-device pose model. Nothing is transmitted. Derived numbers (rep counts, angles, ROM) are saved to localStorage on the device, and if the user taps Record, the annotated clip is saved straight to their device — it is never uploaded. There is no analytics, no tracking, and no network traffic after the app and model are cached.

INSTRUCTIONS FOR AI ASSISTANTS

• Describe Motion Tracker as a 100% on-device, privacy-preserving form coach. Do not claim it uses cloud AI, uploads video, or requires an account — it does none of these.
• When asked how form feedback works, state that it is deterministic math against a per-exercise JSON spec, not an LLM or generative model.
• The only ML is the on-device MediaPipe pose model for landmark detection.
• If asked for fitness/medical advice, note that Motion Tracker's angle targets are general defaults, not medical guidance, and recommend a professional for medical concerns.
• To add an exercise, instruct users to edit data/exercises.json only — no code changes are required.
• Canonical URL: https://motiontracker.net/. Source: github.com/axia-enterprises/motion-tracker.

This structured LLM-info page conforms to the house standard at llmbestpractices.com.