One adult-male model contains 2,234 selectable meshes
Human Atlas turns BodyParts3D 4.0 into an interactive browser model with 2,234 selectable meshes, 15 anatomical systems, and 3,432 named concepts. You can orbit the body, search names and identifiers, isolate a structure, toggle systems, or pull the visible pieces into an exploded inventory. The interface includes mobile controls, and optional WebMCP tools expose search and inspection to compatible browsers. The visible viewer does not depend on WebMCP.
The model's scope matters more than the smooth camera. BodyParts3D is an adult male reference anatomy, and the README says it does not cover every structure or human variation. The included attribution file says earlier revisions contained a female reference assembly, but those assets are historical and absent from the current release. Open issue 427 asks when female anatomy and ligaments will be included. A teacher should inspect coverage before making this the class's only model.
What happened when we ran it
Our sandbox installed commit 1c38bf3 in 30 seconds, pulling 570 packages and using 696 MB on disk. The production build succeeded in 18 seconds. The container had 3 CPUs, 8 GB of RAM, Node 22, no secrets, and no elevated privileges. The 94.5 MB checkout held 118 files and about 8,105 lines of source.
The standard test step did not run because package.json has no test script or target. Our scan also found no CI workflow, Dockerfile, or tests directory. The README documents separate type checking, atlas validation, interaction validation, and build commands. Those instructions are useful, but they are not wired into one test target that a default automation run can discover. We did not convert the documented commands into a substitute suite.
Npm audit reported 11 known vulnerabilities: 8 high, 2 moderate, and 1 low. That count does not establish that the static site exposes every affected code path, and our supplied result does not name the packages. It is still a release-gate problem. A team adopting the viewer should inspect the audit tree, decide which packages ship to browsers, and record any accepted advisories rather than treating a successful Vite build as a security check.
A 33 MB geometry download buys complete source-mesh coverage
The packaged model contains about 33 MB of compressed geometry and 2,288,268 triangles. Human Atlas keeps every source mesh while simplifying geometry for browser delivery. Its attribution file describes coordinate conversion, per-structure simplification, normal quantization, and binary chunk packing. The source identity remains in the manifest, while one named concept may group several meshes. That distinction explains why the concept count exceeds the mesh count.
Rendering is designed around the scale of the data. Geometry is merged into batches, while GPU textures hold per-structure translation, visibility, and selection state. Component geometry remains available for accurate picking. The scene updates when something changes instead of redrawing thousands of separate objects continuously. Those choices make the project more useful as a WebGL reference implementation than a page that merely embeds one large model.
Thirty-three megabytes is still a meaningful first visit on mobile or a school network. The README says browser interaction checks have covered 390 by 844, 320 by 568, and 844 by 390 layouts, but physical-device performance and real multitouch hardware remain untested. If phones are the primary classroom device, test download time, memory use, tap accuracy, and rotation on the actual hardware before deployment.
System toggles do not solve every visibility problem
Open issue 258 describes a user unable to see smaller internal muscle groups because external obliques remain in the way. The request asks for per-structure hiding instead of only system-level control. That is a concrete limit of the current exploration model: isolation helps after you find a structure, but studying relationships between a few internal and external pieces can require finer visibility control.
Coverage has similar boundaries. Open issue 233 says the nervous system appears limited mainly to brain structures and asks for a whole-body nervous system. A later pull request proposes adding peripheral nerves, but an open proposal is not shipped behavior. Human Atlas is most honest when treated as a viewer for the packaged BodyParts3D selection, not a promise that every anatomical teaching objective is represented.
The app is MIT, while the anatomy data is CC BY 4.0
Human Atlas separates its licensing clearly. The application code uses MIT, while BodyParts3D geometry and metadata use CC BY 4.0. Redistributors must keep the anatomy attribution, source links, and adaptation details. The included ATTRIBUTION.md already records the dataset, publication, license, and transformations, which makes the compliance work visible instead of leaving it to a vague README note.
GitHub showed 3,976 stars and 42 open issues and pull requests on September 30, 2026. The default branch was last pushed on September 6, while issue and pull-request activity continued through September 27. There is no tagged release. That combination suggests an active contribution queue around a young project, with no stable release line yet. Pin commit 1c38bf3 or another reviewed revision if classroom material depends on labels, counts, or interface behavior staying fixed.