Analyse

Crux Coach runs Google's MediaPipe Pose model in your browser. It finds 33 body points (shoulders, elbows, wrists, hips, knees, ankles, toes) in about 10 frames per second of your video. It follows the climber with a zoomed-in crop, so far-away climbers are still big enough to track. It also measures camera movement from the rock around them. Each frame is checked twice (normal and mirrored). Legs hidden behind your body are estimated from the frames around them and your body proportions, and they're drawn faded in the replay. From those points it measures:

• Footwork: how often your feet move compared with your hands, and whether they stay put once placed.
• Arm efficiency: elbow angle when you're holding still, and whether upward movement comes from straightening your knees or bending your elbows.
• Body position: how far your centre of mass sits to the side of your feet, and how often you turn your hips into the wall.
• Flow: how direct your body's path is, sudden jolts, hesitations and grip readjustments.

It also works out how the climb ended (top-out, matched finish, lower-off or fall). Coming off within about a second of reaching a hold counts as a fall. For every fall it replays the last seconds and looks for the cause: missed catch, late dead-point, over-reach, feet cutting, hand or foot slip, barn door, lock-off, stalling or pump. Every finding shows how sure the app is.

Every half second is also labelled with what your body and each hand and foot were doing, following a research-based framework for climbing video. Anything one camera can't show (grip type, how close your hips are to the wall, whether a foot is weighted) is marked "unknown" instead of guessed. You can export these labels.

The coaches. Magnus, Alex and Eric are the three analysis levels (the names are a playful nod to well-known climbing coaches and YouTubers, not the real people). They're the same rules-based analysis, not different AIs. Magnus uses the fastest, lightest pose model and shows a short report. Alex uses the standard model with a mirrored second check and shows the main report, with the technical extras under "More detail". Eric uses the largest model and shows everything. Outdoor or indoor changes how hard the app searches for a small, far-away climber, and swaps rock-specific tips (brushing holds, cool conditions) for gym ones (polished holds, chalk build-up, loose holds).

Each measurement is scored 0–100 against coaching rules. It's then turned into a move-by-move review, a start/middle/top comparison, left vs. right, and a top-3 plan with drills and targets for your next session.

• It sees a flat 2D picture. It can't measure how close your hips are to the wall, the wall angle, or where the holds are.
• The pose model wasn't trained specifically on climbers. Unusual positions, occlusion or poor light can confuse it.
• Camera pans are compensated for by tracking the rock in the background, and gradual changes in your size on screen (climbing away from the camera) are rescaled. Sudden zoom jumps, and plain sky with no rock to track, are still hard, so keep the camera steady there.
• Very far away (the climber is only a few dozen pixels tall) gives rough hand and foot measurements. Record in 4K or zoom in.
• The scores are guidance, not a verdict. Use them to spot trends over many sessions.
• It can't see the holds, so it can't know whether the last hold you held was the real finish. That's why you can correct the result. And it's a rules-based analysis of your body movement, not an AI trained on climbing videos, so check its fall explanations against the replay.

Your videos never leave your phone. There are no accounts, no servers and no tracking. Saved sessions contain only the measurements and a small thumbnail.