Spaceforge

A browser-local website builder. A small language model (Gemma 4 E2B by default; larger Gemma and Qwen-Coder tiers available) runs in desktop Chrome via WebGPU and writes your site's HTML/CSS/JS files into localStorage. Preview the site in a sandboxed iframe, edit any file in Monaco, download individual files or the whole site as a .zip.

No backend for the model itself. One tiny Vercel Edge function (api/photo.ts) proxies Unsplash server-side so the API key never reaches the client bundle. After the first multi-GB model download the LLM runs entirely offline.

Requirements

• Desktop Chrome 134+ with WebGPU (shader-f16 feature).
• ~3 GB free RAM for the default model; more for larger tiers.
• ~2 GB of free disk space for the first-time model cache.

How it works

1. transformers.js loads an ONNX-quantized Gemma/Qwen model via WebGPU.
2. Your chat messages are streamed through the model. The output uses a simple ===FILE: path=== / ===END=== protocol so the model emits whole files, not diffs.
3. A streaming parser writes each completed file into localStorage.
4. The site is authored in 11ty-style conventions: Markdown pages (index.md, about.md) with YAML front matter, Nunjucks layouts and partials (_layout.njk, _header.njk, _footer.njk). The preview renders .md through its layout on the fly; the zip export bakes everything to .html.
5. Three tabs on the right:
   • Preview — sandboxed iframe; link clicks are intercepted and routed back through the parent.
   • Source — file tree + Monaco editor.
   • Template — swap between "Custom (AI-generated)" (default; the model writes its own layout + styles) and pre-existing template bundles that override those files. Overlay is non-destructive — switching back restores the generated files.

Scripts

npm install
npm run dev        # Next.js dev server on :3000
npm run test       # unit tests (vitest)
npm run typecheck  # tsc --noEmit
npm run build      # Next.js production build
npm run start      # serve the production build locally

Local dev environment

Spaceforge runs end-to-end locally with no cloud dependencies. One-shot:

npm install
npm run dev:local:setup   # docker compose up -d + db:migrate + db:seed:dev
npm run dev:local         # Next.js on :3000 with AUTH_DRIVER=dev + FsBlobDriver

Open http://localhost:3000 — you're auto-signed in as the seeded dev user and land on the dashboard. Create a site, edit it, etc.

Driver toggles (set via env, defaults shown):

• AUTH_DRIVER=dev — fixed fake user, no Clerk required. Set to clerk to use real auth (needs keys in .env.local).
• BLOB_DRIVER=fs — writes to .spaceforge-local/blob/. No network, no Docker, no token. Set to vercel for real Vercel Blob in preview/production.

Useful sub-scripts:

• docker compose up -d — just start Postgres.
• npm run db:migrate — apply Drizzle migrations.
• npm run db:seed:dev — ensure the dev user/team rows exist.
• npm run db:studio — browse the DB in Drizzle Studio.

Environment variables

See .env.example for the canonical list. At minimum for the photo proxy:

• UNSPLASH_ACCESS_KEY — your Unsplash app's Access Key (not the Secret Key).

Phase 1+ also needs DATABASE_URL, NEXT_PUBLIC_CLERK_PUBLISHABLE_KEY, CLERK_SECRET_KEY, CLERK_WEBHOOK_SECRET.

Deploy

Next.js on Vercel. COOP/COEP headers (required for WebGPU) are set in next.config.ts. Route handlers under app/api/ are auto-deployed as Vercel Functions (Fluid Compute).