SuperSplat editor and splat-transform
Two pinned tools from PlayCanvas (MIT, ADR-2177), gated by
[skills.spatial_and_3d] supersplat in agentbox.toml:
| Tool | Where | Use |
|---|---|---|
| SuperSplat 3.6.0 | http://agentbox:8795/editor/ (static, served by [program:supersplat]) |
Look at a splat, select and delete splats, crop, render, export |
splat-transform 3.10.3 |
on PATH | Headless conversion, compression, filtering, merging |
Before you start
Probe the server; a gate that is on but not yet built into the image shows up here:
curl -s -o /dev/null -w '%{http_code}\n' http://127.0.0.1:8795/editor/ # 200 = served
supervisorctl status supersplat
command -v splat-transform
No answer means the image predates ADR-2177 or the gate is off: say so and stop,
or fall back to splat-transform from a scratch npm install @playcanvas/splat-transform@3.10.3.
Opening a splat
The server's root is $WORKSPACE/splats/share. Copy or convert a splat into its own
folder there, then open it. The editor and the file share one origin, so no CORS.
mkdir -p "$WORKSPACE/splats/share/workshop"
splat-transform scene.ply "$WORKSPACE/splats/share/workshop/scene.compressed.ply"
Open http://agentbox:8795/editor/?load=/workshop/scene.compressed.ply in the browser
sidecar (browser skill). Repeat load= for several files.
WebGPU only. SuperSplat 3.6.0 creates a WebGPU device and has no WebGL2 fallback,
so the page must be a secure context. http://agentbox:8795 is on the browser
sidecar's TREAT_AS_SECURE list. That list is read when the sidecar starts, so on
an older sidecar forward the port to localhost inside it and open
http://localhost:8795/editor/... instead:
docker exec -d browsercontainer socat TCP-LISTEN:8795,bind=127.0.0.1,fork,reuseaddr TCP:agentbox:8795
Do not recreate the browser sidecar to pick the list up without asking: it locks the owner's Podkey until they unlock it again over VNC.
Driving the editor
The page exposes window.scene; window.scene.events is the editor's own event bus.
Wait for the import, then check you are on hardware WebGPU before trusting anything:
async () => {
const sleep = ms => new Promise(r => setTimeout(r, ms));
for (let i = 0; i < 60 && !window.scene?.events; i++) await sleep(1000);
const ev = window.scene.events;
let splats = [];
for (let i = 0; i < 120 && !splats.length; i++) { splats = ev.invoke('scene.splats') || []; await sleep(1000); }
const a = window.scene.graphicsDevice?.gpuAdapter;
return { secure: isSecureContext, adapter: a?.info?.vendor,
fallback: a?.info?.isFallbackAdapter ?? a?.isFallbackAdapter,
files: splats.map(s => ({ name: s.name, numSplats: s.numSplats })) };
}
fallback: true or a vendor of google means software rendering: report it, do not
judge quality from it. The verified calls and an edit recipe are in
the event-bus reference. That surface is SuperSplat's
internal API, pinned to 3.6.0: re-check it against src/ after a version bump.
splat-transform
Formats in: .ply .compressed.ply .sog .spz .ksplat .splat .lcc. Out adds .glb .csv .html.
splat-transform in.ply out.compressed.ply # ~4x smaller, CPU, seconds
splat-transform in.ply out.spz # Niantic SPZ
splat-transform -g cpu in.ply out.sog # PlayCanvas SOG, smallest; -g cpu works
splat-transform --help # transforms, filters, merging
Measured on the 1.04 M-splat workshop scene (71 MB PLY): compressed PLY 16.2 MB,
SPZ 15.7 MB, SOG 11.2 MB, each in under 4 s on CPU. The CLI loads the Dawn WebGPU
addon on every command, but this container exposes no WebGPU adapter
(--list-gpus finds none), so pass -g cpu. The GPU-only features
(--filter-cluster, --filter-floaters, .voxel.json) need a host with one.
Limits
- Exports from the editor download inside the browser sidecar, not to this container.
Prefer
splat-transformfor files you need back here, and the editor for looking and for selection edits you then reproduce headlessly. - The share root is readable by every container on the docker network. Put only
splats there that may be seen; datasets stay in
$WORKSPACE/splats/<name>/.
