Uruk composes entire sites from rows in Postgres — element trees rendered to HTML,
cached until the data changes, edited by people and AI agents over the same API.
No templates to commit. No build to run. No deploy to wait for.
sh — get running
$ git clone https://github.com/Galaktis/uruk && cd uruk$ docker compose up# http://localhost:8000 → this site, served from rows
the stack is a ziggurat — state at the base, HTML at the summit
From tablet to page
Three moves, same as 3200 BC
A scribe pressed records into clay; a renderer presses rows into HTML. The pipeline is short enough to hold in your head.
01 · WRITE
Rows, not files
Every block on this page is an Element row — a kind, JSON parameters, content, and a parent. Pages are rows too, arranged in a URL tree.
{
"kind": "heading",
"content": "A website is data",
"parameters": { "level": 1 },
"parent_id": "0198…root"
}
02 · RENDER
The tree becomes HTML
A small library of committed renderers walks the element tree. Text is escaped by default; there are no per-site templates anywhere.
# the only committed "template"@register("heading")
defheading(el, children, ctx):
return format_html(
"<h{0}>{1}</h{0}>",
el.parameters["level"], el.content)
03 · CACHE
Cached until a row changes
Each render records which elements it touched. Saving one element evicts exactly the pages that used it — a shared header invalidates everywhere at once.
element.save()
└─ PageDependency reverse index └─ evict: / /manifesto …✓ next request re-renders once
What the city provides
One deployment, many sites
A single Uruk serves any number of tenants. Each one is a walled district: its own schema, its own users, its own everything.
A schema per tenant
The Host header picks the tenant; the connection's search_path does the isolation. Users, content, and tokens never share tables across sites.
Exact invalidation
The hard problem, solved with a dependency graph: every render records what it read, so one edit evicts precisely the affected pages — nothing more, nothing stale.
Agents are editors
The first admin client is an MCP server, not a dashboard. An agent authenticates as a tenant user and can do exactly what that user's permissions allow.
Sites that travel
Every row has a UUIDv7 key, so a whole site exports as one JSON archive and restores anywhere without collisions. This very site ships as one.
Assets live in the DB
The CSS styling this page is a row, served at a content-hashed URL with an immutable year-long cache. Even the fonts are rows. Few files, full speed.
No commit-to-create
A new site is an API call, not a repository. The only committed code is the generic renderer library every tenant shares — everything else is rows.
Founding a city
Running in one command
Docker is the only requirement. On first start the app container
migrates every schema and raises the default tenant — the site you are reading now —
next to a remote MCP endpoint. Then hand an agent a token, let it build, review the
draft in the Studio, and publish.
$ git clone https://github.com/Galaktis/uruk && cd uruk$ docker compose up -dapp migrating public + tenant schemas, seeding the default site…app Listening at: http://0.0.0.0:8000mcp Listening at: http://0.0.0.0:8001 (/mcp)# any hostname you point at it now serves this site