SPEC v7 · browser-native compute substrate

Compose, compile and train AI in a browser tab.

Scellis turns the browser into a compute substrate: draw a workflow, lower it to WebGPU, and train models locally — with your data never leaving the device.

[ 0 install · 0 deps · 0 uploads ]

Fig. 1 — compute field, rendered live on your GPU

§1Capabilities

The whole platform, on one screen

A small, stable engine — every capability delivered as content from the database.

ParameterValueNote
InstallnoneOpen a tab and go — no CUDA, no Python runtime
ComputeWebGPURuns on the GPU already in your machine
Dataon-deviceEgress is a declared, consent-gated effect
Reproduceby URLContent-addressed artifacts + fixed seeds
Extenduser = builtinAuthor blocks, kernels, viewers, models
Trainon WebGPUExplicit budgets, autodiff, oracle-checked numerics
§2How it compiles

A drawn graph, lowered to GPU code

Your workflow descends a ladder of representations — and runs where you drew it.

L4Workflowwhat you draw
scellis.com/app/webDatasetdataLayersmodelTraintrainerEvaluatemetricsShareexportruntraining lossloss 0.04 · epoch 7/8
L1Canonical IRautodiff · memory plan
; canonical IR
%x = load "input"
%w = param [768, 256]
%h = matmul %x %w
%y = relu %h
L0WebGPUdispatched on your GPU
// webgpu dispatch
@compute @workgroup_size(64)
fn main(id) {
  // 256 invocations
}

Fig. 2 — Lowering pipeline · Workflow → IR → WebGPU

§3Local-first

Your data never leaves the device

Computation is local; egress is a declared, consent-gated effect — never the default.

your device
[B, 768][B, 256][B, 10]

Nothing crosses the boundary unless you grant a declared effect.

Fig. 3 — Data path, bounded to the device

§4By the numbers

Measured, not promised

0
installs to get started
100%
on-device by default
1
browser tab is the runtime
capability, via content
§5Extensibility

Built-ins are just the starting set

Everything is content delivered from the database. User-authored entries are structurally identical to the built-ins.

ItemPartSource
01blocksseed
02kernels (wasm + webgpu)seed
03your blockuser/
04viewersseed
05your modeluser/
06datasetsuser/

Built-in and user rows are drawn identically — by design: one registry, one validation, one path.

Abstract

Scellis treats computation as a composable, inspectable, and shareable artifact. The core stays small and stable; capability grows, without bound, through content. WebGPU replaces CUDA, the database replaces the filesystem, and every run is reproducible — bit-for-bit — from a link.

epoch 8/8 · loss 0.04 · converged

Open a tab and do science.

No install, no upload, no data center. Your first workflow is a click away.

End of spec · revision 7