A lightweight Node.js CMS where an LLM acts as the semantic translation layer between a block editor and accessible, mobile-first HTML.
Non-technical editors arrange content blocks. On save, the block JSON is passed to an LLM which returns clean, semantic HTML — with correct heading hierarchy, meaningful alt text, and a mobile-first layout. The editor never touches HTML, CSS, or media queries.
cp .env.example .env
# Edit .env — set ADMIN_PASSWORD, SESSION_SECRET, and your LLM API key
npm install
npm start
# Visit http://localhost:3000/adminAll configuration is via environment variables. See .env.example for the full list.
| Adapter | Env var | Default model |
|---|---|---|
anthropic |
LLM_API_KEY |
claude-sonnet-4-6 |
openai |
LLM_API_KEY |
gpt-4o |
ollama |
(none) | llama3 |
Set LLM_ADAPTER to switch. Ollama runs locally with no API key required.
| Adapter | Notes |
|---|---|
local |
Dev only — stores files in public/uploads/ |
r2 |
Cloudflare R2 (recommended for production) |
s3 |
AWS S3 |
backblaze |
Backblaze B2 |
The semantic HTML prompt and CSS audit prompt are plain text files in prompts/. Edit them to tune the LLM output for your site.
heading— h1, h2, or h3paragraph— body text with inline formattingimage— single image with caption and flow (left/right/full/center)gallery— grid or strip of imagesdivider— section breakfile— downloadable attachmentresource-list— collection of downloadable filescontact-form— first-class contact form, wired to SMTP
# Install PM2
npm install -g pm2
# Start
pm2 start ecosystem.config.cjs --env production
pm2 save
pm2 startupNginx config and full deployment instructions are in docs/technical-spec.md.
See docs/cms-spec.md for the full project specification.
Stack: Node.js 20+, Express, SQLite (better-sqlite3), vanilla JS + Web Components (no framework, no build step).
RAM footprint: ~80–120MB idle. Runs on a $5 VPS.