He Put a Fitness Coach in the Browser—But It Won't Score Your Moves

Qian Jiang posted a demo on X: following along with a workout video at home, with the camera pointed at the person, joints marked on screen in real time, and movement cues displayed. The project is called Form Mirror, credited on the page as a Local Movement Lab prototype. On the left is the tutorial video; on the right is the user's feed, with joints dotted and a line of text on screen: Continuous cloud coaching is on.

The pain point of exercising at home is very concrete. The instructor in the video has perfect form, but you off camera can hardly see whether your shoulders are shrugging or your elbows are locked. A user, Sophie Bennett, said in a reply that real-time feedback can make home workouts more purposeful, especially when it is hard to judge whether your form is correct. That sentence makes the need clear.

Qian Jiang added that AI-generated movement explanations and muscle diagrams overlaid on video screenshots are very useful.

The image is labeled 'Muscle group illustration—not actual activation detection' and gives a 43% completion rate. This detail carries more information than the words 'AI coach': it knows what it is doing, and it also knows what it cannot do.

The trial page lays out the features. You can use the sample workouts included on the page, paste a YouTube link, or upload a video file. As you follow the video and move, the browser accesses the camera and estimates the angles of the left and right elbows and the arm raise angle. The page suggests placing the camera at chest height, keeping shoulders, elbows, and wrists visible for arm movements, facing the camera, and avoiding strong backlighting. The angles are 2D estimates; side angles, occlusion, and perspective all affect the results. The page does not provide a form score, and it also states that the timeline and real-world accuracy have not been verified.

Not scoring is smart restraint. Fitness apps easily slide toward scores, rankings, and streak check-ins, eventually becoming another source of anxiety. Form Mirror only gives cues, breaking 'is my movement correct?' into joint angles and voice feedback. The page also states that the movement illustrations and annotations come from the original research and were generated with AI assistance, are not muscle activation measurements, and have not been reviewed by professional anatomists. This active exposure of limitations is more credible than pretending to be professional.

In the replies, someone asked whether it is built on MediaPipe. The author didn't take the bait and only replied, 'ship.' Someone joked that finally there's an AI that makes humans do the heavy lifting. Someone asked what video model this is. Another person said they had just built exactly the same thing. The Archer mentioned that the first-place project at the HackTN26 hackathon used a similar approach. These reactions show one thing: using a browser camera plus pose estimation for movement feedback has lowered the technical barrier enough that individual developers can quickly assemble a prototype. The hard part is not 'can it recognize joints,' but movement segmentation, latency, viewpoint robustness, and feedback wording that doesn't annoy people.

Qian Jiang provided a trial link: ai-workout-companion-phi.vercel.app. On Vercel, you can open it and try it.

There is another reply worth noting for people building health AI. Quentin Caron posted a GitHub link to Open Health Sync, and Qian Jiang replied with thanks. Open Health Sync is an open-source, self-hosted health data aggregation layer that supports 18 sources including Apple Health, Withings, Oura, WHOOP, Fitbit, Health Connect, Dexcom, and OMRON. It is a single Go binary plus Postgres and can run on a Raspberry Pi. Outputs include a REST API, FHIR R4, and a read-only MCP server for AI agents to read data. In other words, products like Form Mirror solve the front end of 'movement feedback,' while Open Health Sync fills in the back end of 'where health data comes from, how to normalize it, and how to safely let models read it.' The two are not at the same layer, but viewed on the same timeline, AI tools for home fitness are moving from demos to parts that can be assembled.

Project link: Open Health Sync

Form Mirror is not yet a mature product. The page itself says it is a prototype, and the timeline and accuracy have not been verified. But it gets one thing right: first let the camera see the movement, then let feedback happen. It doesn't score, it doesn't boast, and it writes its limitations where users can see them. For people practicing movements at home, this is more practical than a flashy AI coach.