Product · Flagship
PoseTracker: why we built it
PoseTracker is made by Movelytics. It gives an app the ability to understand a body in motion: it runs on the device and returns keypoints, joint angles, rep counts, a form score, jump height and flexibility measures. We built it because the hard part was never the model. It was making a model work inside a real app, on every phone.
Last updated:
01Problem
Open pose models exist. Shipping one is the problem.
Several good pose estimation models are public. A developer can download one in an afternoon and see a skeleton drawn over a webcam image. That demo is the easy part.
The rest is everything around the model, and it costs months:
- The camera pipeline. Capture, orientation, mirroring, pixel formats, resizing to the model's input. It is different on iOS, on Android and on the web, and a large share of "pose bugs" are really camera bugs.
- Hardware acceleration. The fast path changes with the device, the operating system version and the driver. When it fails, it often fails silently: the frame rate drops and nobody knows why.
- Drift. Operating systems and machine-learning libraries change all the time. An update breaks something that worked. The integration is never finished.
- The layer above the model. A model returns points. A product needs repetitions, angles, a form score, a comparison with a reference movement. That logic has to be written, tested and maintained on each platform.
02Why we built it
One integration instead of three pipelines
PoseTracker removes that work. A team adds movement understanding to its app in minutes, and the pipeline, the acceleration, the stabilisation across iOS, Android and the web, and the logic above the model are ours to maintain, not theirs.
Two commitments shaped the design from the first version.
- The video never leaves the phone
- Inference runs on the user's device. There is no upload for analysis, so there is no server-side video for anyone to store.
- The host app owns the data
- PoseTracker does not store end users' analysis history. The app that integrates it decides what to keep, in its own system.
03Business model
Three ways to work with PoseTracker, and who owns what
PoseTracker has always been built in two layers. The pose model at the bottom has historically been an open-source one, such as MoveNet. On top of it we build our own movement-recognition engines: the logic that turns keypoints into repetitions, angles, scores and comparisons. That second layer is our work.
Many of our clients do not want to rent this technology. They want it among their own assets: the intellectual property and the whole motion-tracking stack. Others want the most capable model and are happy to license it. So the business model follows one question: which pose model do you choose?
| Self-serve API | Custom build, open-source model | Custom build, PoseTracker Model | |
|---|---|---|---|
| What it is | The PoseTracker SDK and API, integrated by your developers | A PoseTracker built for you by the lab, on an open-source pose model | A PoseTracker built for you by the lab, on our proprietary model |
| Pose model | Open-source model, such as MoveNet | Open-source model, such as MoveNet | PoseTracker Model: optimised for mobile, 39 keypoints instead of 17 |
| Who owns the technology | Movelytics. You use it as a service | You. The full intellectual property of the build is delivered to you | You own the custom development; the model stays ours and is licensed to you |
| How you pay | Metered usage (API calls), with a free tier | Custom development | Custom development, plus an annual licence for the model |
In short: custom development, with or without the full intellectual property delivered to you, depending on the model you choose. An open-source model means you can own everything. Our model means more keypoints and a model tuned for the phones your users hold, under an annual licence.
We built our own model for a second reason, which matters to every API user: stability. Depending on a public model that someone else may stop maintaining is a risk for a product. Owning the model lets us keep the API stable and relevant over time. The details are on the PoseTracker Model page; to discuss a custom build, see work with the lab.
04How it works
The app asks for a movement. PoseTracker returns events.
- The app opens PoseTracker with its token and names the movement to track.
- PoseTracker runs pose estimation on the device, frame after frame, from the camera.
- It computes the results: keypoints, joint angles, repetitions, a form score, progress.
- It sends structured events back to the app, which updates its interface and stores what it wants.
Push-up · elbow angle and body lineIllustration. Authored pose sequence, not live inference.
It works on React Native, iOS, Android and the web. The full pipeline, from camera frame to rep count, is described on our technology page.
05What you get
Outputs an app can use directly
- Keypoints
- Named points on the body with their position and confidence, for each frame.
- Joint angles
- Angles at the knee, hip, elbow, shoulder and more, computed from the keypoints.
- Rep counts
- Repetitions detected from the movement itself, exercise by exercise.
- Form score
- A score for how the movement was performed.
- Jump height and flexibility
- Specialised measures for jumps and for flexibility tests.
- Comparison with a reference
- The user's movement compared with a reference movement, and scored.
Pricing, plans and integration guides live on the product site. This page tells the story; posetracker.com is where developers start.
06Proof
Who uses it
- 1,250+ registered developers as of 17 November 2026.
- In production in apps including Flexifit, Kobi, How2Jump and MoveBreak.
- A free tier of 200 API calls per month, so a team can test on real phones before any commitment.
PoseTracker is also where our own model arrives first. The PoseTracker Model, in private beta, is designed to bring real-time tracking with 39 keypoints to low-end phones.
07Data and privacy
What PoseTracker knows about your users: almost nothing
The camera image is analysed on the phone. Movelytics receives a usage count, the number of API calls, which is the basis of billing. We do not receive the video, the keypoints or the results, and we do not keep an analysis history for end users.
This is a design choice, not a setting. It is described in full in our data commitments. The developer's guide to this topic is PoseTracker's page on on-device pose estimation.
08FAQ
Questions about PoseTracker and Movelytics
Who makes PoseTracker?
PoseTracker is made by Movelytics SAS, a French applied-AI lab based in Bidart, in the Basque Country. The same team trains the PoseTracker Model and builds MoveBreak.
Does PoseTracker upload video to a server?
No. Pose estimation runs on the user's device. The video is not uploaded for inference, and Movelytics does not store end users' videos, keypoints or analysis history.
Which platforms does PoseTracker support?
React Native, iOS, Android and the web. Integration guides are in the PoseTracker documentation.
Is there a free plan?
Yes. PoseTracker has a free tier of 200 API calls per month. Plans and pricing are on posetracker.com.
Can Movelytics adapt PoseTracker to a movement it does not cover?
Yes. The lab builds custom keypoints, fine-tuned pose models and custom movement logic for clients. See work with the lab.
Can we own the intellectual property of our PoseTracker integration?
Yes. With a custom build on an open-source pose model such as MoveNet, the full intellectual property of the development is delivered to you. With the proprietary PoseTracker Model, you own the custom development and the model is licensed to you annually.
09Go further
Start
Add movement understanding to your app.
Test it on real phones with the free tier, or talk to the team that builds it.