Fundamentals / Overview
Version 1.2.0
Carapace is a content management system with no CMS server. This page covers the core idea before anything practical — what it actually is, and who it's built for.
The short version
A typical headless CMS is three pieces of infrastructure you run and maintain: a database, an API layer, and an admin UI. Carapace has none of them. Git is the database. An MCP-connected AI agent is the admin UI. And there's no API layer at all — the site's build reads content straight from the repository. A visitor's request never touches anything you'd call "the CMS."
Three tiers, split by what can go wrong
Everything in Carapace falls into one of three tiers, and the split is deliberate: it's based on risk, not on how a typical design system is organized.
- Content — the HTML you write. Inert by construction; it can never execute anything, so an agent can write it directly and safely.
- Style — colors, spacing, a handful of layout values. Editable as structured data within a fixed, validated vocabulary — never open-ended CSS.
- Execution — real code: templates, components, the build itself. Always human-written, always reviewed. An agent never writes this tier directly, on purpose, permanently.
Two roles map onto the first two tiers — Writer and Designer — with tool access to match. The third tier isn't a role at all; it's ordinary code, reviewed the normal way.
Who this is for
Developer-led teams and solo builders whose content is shaped like a site — a blog, docs, marketing pages — who want to write by talking to an agent, and who'd rather their content lived as reviewable git history than as rows in someone else's database.
It's not for enterprise marketing orgs running complex approval chains, newsroom-scale concurrent publishing, or anything needing real localization or a multi-channel content hub. That's a permanent design boundary, not a gap on a roadmap.
Ready to set it up? Continue to Getting Started.