Open-source human-atlas makes 2,234 BodyParts3D meshes explorable in the browser

A new open-source project called human-atlas, which surfaced on Hacker News's front page on 9 September 2026, turns a full human anatomy reference into an interactive 3D scene that runs entirely in the browser. According to the project's GitHub repository, the viewer breaks the BodyParts3D adult male reference into 2,234 individually selectable meshes, organised under 15 anatomical systems.

The application is built with React, Three.js, and shadcn/ui. It requires no API keys or accounts, and the repository ships browser-ready geometry, so the whole thing can be built and served as a static site.

What the viewer can do

Users can orbit, zoom, and pick structures directly on the model, or switch to an exploded view that spreads the visible pieces into a spaced inventory. Individual system layers can be toggled, and preset views for the skeleton and the organs offer quicker starting points. A search index covers 3,432 named anatomical concepts, and isolating a structure brings up a detail panel with its information. The repository notes that phone-sized interfaces were also a design target, with compact controls and detail panels that stay out of the way of the exploded layout.

The anatomy data underneath

The geometry comes from BodyParts3D 4.0, an adult male reference anatomy licensed under CC BY 4.0. The README is frank about the dataset's limits: it does not represent every human structure or variation, and the project positions itself as an educational explorer rather than a diagnostic or surgical tool.

One subtlety the documentation calls out is that individual meshes are distinct from named concepts, which may group several meshes under a single name. Geometry is simplified for browser performance while retaining every source mesh: the packaged model contains 2,288,268 triangles and downloads as roughly 33 MB of compressed geometry, with simplification held to a 0.2 percent relative error limit per structure. Full credits and adaptation details are kept in an ATTRIBUTION.md file.

Keeping thousands of meshes interactive

The rendering approach is the technically interesting part. Geometry is merged into batches, and per-structure GPU textures drive translation, visibility, and selection, while component geometry handles accurate picking. Exploded layouts pack only the pieces that are actually visible, and the scene re-renders when it changes rather than continuously — so, according to the repository, orbit controls stay responsive without thousands of separate draw calls.

There is also an optional set of WebMCP tools that expose anatomy search and inspection to compatible browsers; the visible interface works without them.

Validation gets unusual attention for a hobby-scale project. Scripts check mesh buffers, names and concept membership, non-overlapping exploded layouts at desktop and mobile aspect ratios, search and inspection behaviour, and tap-versus-drag handling. Browser-based tests have exercised selection, system controls, search, isolation, and rotation at 390×844, 320×568, and 844×390 layouts. The README concedes that real multitouch hardware and physical-device performance have not been tested.

Running, rebuilding, and licensing

Running it locally needs Node.js 22.13 or newer. A build produces static output in the dist directory, which can be deployed to Vercel as a Vite project using the included configuration, or to any static host.

Rebuilding the geometry from scratch is optional: developers can fetch the official BodyParts3D OBJ archive and English metadata tables, then run the repository's conversion, optimisation, and compression scripts. Licenses are deliberately separated — the application code is MIT, while the anatomy data keeps its CC BY 4.0 terms, and attribution must be preserved when redistributing it.

Why it matters

BodyParts3D has been openly licensed for years, but exploring it usually meant downloading archives and wrestling with desktop 3D software. Packaging it as a self-contained static web app makes a full anatomical reference explorable on a phone, with no accounts or infrastructure — a meaningful accessibility win for students and educators.

For developers, the batching-plus-GPU-texture approach is a reusable pattern for anyone who needs to render and pick thousands of objects in WebGL without collapsing the frame rate. And the project sets a good example on process: clean separation of code and data licenses, honest disclosure of untested hardware scenarios, and an explicit statement that it is educational, not clinical. The single male reference and simplified geometry limit how far it can go, but as a free, inspectable teaching tool, it lowers the barrier to exploring human anatomy about as far as it can go.