[ DECLARE ]
Intent-driven,
not click-driven
Declare the state you want. Watch it reconcile.
awaiting intent
3 of 8 sampled items drift from this intent
desired
5/8
[ THE CONTENT OPERATING SYSTEM · TOTALITY EDITION ]
Your content,
operated by AI.
LumiBase is the Content Operating System — agents do the operational work, you set the intent and hold the veto.
[ 00 / ORIENTATION ]
A CMS stores.
An OS operates.
A CMS is where content sits and waits for someone to touch it. A Content OS is what runs it: you declare the state your content should be in, and a governed control loop keeps it there.
Traditional CMSLumiBase Content OS
Unit of work
CMSOne operation at a time — a person opens an entry and edits it.
Content OSA declared intent: the state your content should be in.
Who operates
CMSPeople. The tool waits to be told what to do.
Content OSGoverned agents pursue the goal; you set intent and taste.
When it happens
CMSWhen someone remembers, or when a ticket is filed.
Content OSContinuously. Drift is detected and reconciled as it appears.
Trust model
CMSRole permissions — access is granted once, all or nothing.
Content OSAutonomy is earned per capability, L0 shadow → L4 autopilot.
Safety net
CMSUndo, after the change is already live.
Content OSConstitution, veto window and kill switch — before it publishes.
What is recorded
CMSAn activity log: who clicked what, and when.
Content OSProvenance: which agent, under which intent, on what evidence.
[ 01 / AI HARNESS ]
AI Harness
Agents operate your content. You set the intent — and hold the veto.
Goals, runs & real skills
The first-class API is a goal, not a click. Agents pursue it through governed skills — every run inspectable, replayable, and attributed.
✓Draft 12 release notesL2 · auto
⏸Rewrite pricing pageawaiting approval
✓Translate docs → vi + enL3 · veto window
•Reconcile SEO titlesrunning
Trust gradient · L0–L4
Autonomy is earned, not granted. Promotion needs a human and moves one level; demotion on incident is automatic. Play the ledger below.
L0 → L4 · earned
Human-in-the-loop
schema:write and delete always route through approval — at every autonomy level, with no exceptions.
Provenance-first
Every revision records agent, run, model, references, and approver — served on the Delivery API with ?provenance=true.
agent:writer-2drafted body · run #482human:maiapproved · constitution ✓agent:translatorlocalized vi → en[ 01B / EARNED AUTONOMY ]
It only watches.
The agent runs, but its output is just recorded for review. Your content is never touched.
L0 · Shadow
L4
L3
L2
L1
L0
CLIMBING TAKES WEEKS OF CLEAN WORK · FALLING TAKES ONE INCIDENT
[ 02 / CONTENT OS ]
Content OS
Declare the desired state. LumiBase reconciles content toward it, continuously.
Intent-driven,
DECLARE INTENTRECONCILE CONTINUOUSLYEARN AUTONOMYKEEP THE VETO
[ 03 / STUDIO ]
Studio
Mission control — humans supervise the system instead of hand-editing every item.
[ ON DUTY · 1 human · 12 agents ]
You review the exceptions, not the 1,284 items.
exception inbox
12 agents working
products/ban-phim-co-k3writer · rewrote price copy· waiting on youdocs/en/security/user-managementtranslator · touched an auth doc· waiting on youcollections/orders — schemaarchitect · drops a field· waiting on you
cleared without you
44 / 47
No-code Collection Builder
Any content shape, no migration. Drag-drop UI and live JSON schema, in sync both ways.
titlestring
slugstring
authorm2o → users
bodyrichtext · encrypted
Field-level permissions
RBAC down to one field, and AES-GCM on the fields that need it.
Security
Exception inbox
Approvals, veto windows, incidents and the kill switch in one queue.
Mission Control
Realtime collaboration
Presence and live subscriptions over WebSocket — humans and agents on the same item.
[ 04 / RUNTIME ]
Runtime
Edge-native, never locked in — Cloudflare Workers or self-hosted Docker, one abstraction.
Typegen + SDK
Your schema becomes TypeScript types, and a full JS/TS SDK covers auth, items, files, and realtime.
import { createClient } from '@lumibase/sdk' const db = createClient<Schema>(url) const posts = await db.items('posts').read()
Runtime abstraction
@lumibase/runtime swaps cache, storage, DB, search, and queues per deployment target. Business logic never touches a vendor binding.
CF Workers · Docker
MCP server
Your content becomes tools any AI assistant can call — the API and MCP are the front door, the UI is a projection.
⌘
@lumibase/mcp-serverClickHouse CDC + materialized reads
Stream every change out; materialize hot read paths for the edge. Analytics without touching the write path.
Postgres→CDC→ClickHouse
[ 05 / FIELD MANUAL ]
Questions, answered
What is a Content Operating System?
A traditional CMS is a tool humans use to manipulate content one operation at a time. A Content Operating System inverts that: you declare the desired state of your content (its SLOs), and a control loop of governed AI agents continuously reconciles content toward that state — with full provenance and human veto. LumiBase is that runtime.
How is this different from an AI plugin bolted onto a CMS?
Autonomy in LumiBase is earned, not granted. Every (site, agent, capability) sits on a trust ladder from L0 (shadow) to L4 (autopilot). Agents are promoted by passing evaluations and auto-demoted on incidents. Every action is gated by a versioned tenant constitution and recorded with provenance.
Do humans lose control?
No. Humans set intent, taste, and policy, and hold the veto. The L3 veto window stages dangerous changes and commits them only if no human objects within a time budget. A four-scope kill switch can halt agents instantly. Nothing publishes that fails the constitution.
Is it open source?
Yes. The core LumiBase is free forever under the Apache 2.0 license. It runs edge-native on Cloudflare Workers and can also be self-hosted with Docker via a shared runtime abstraction.
[ FINAL TRANSMISSION ]
Run your content like a system.
Declare the desired state. Let governed agents converge it. Keep the veto. Free forever under Apache 2.0.