Hello Concrete community!
I’ve just open-sourced Macareux Atlas Framework, an architecture for securely connecting AI agents to Concrete CMS.
Now you can start building AI assistants for your Concrete CMS sites without having to solve authentication, permissions, or content retrieval from scratch.
Why “Hours, Not Weeks”?
Because Atlas isn’t just code.
The repository includes a detailed architecture specification written using a Spec Driven Development approach.
I used that specification as the starting point to build the AI assistant now running on our own Concrete CMS documentation site.
From specification to a working AI agent integrated with Concrete CMS took only a few hours.
The goal of Atlas is to let both human developers and AI coding assistants understand the architecture before writing a single line of code.
What problem does it solve?
If you’ve ever tried to connect an LLM to a CMS, you’ve probably asked questions like:
- How does the AI know who’s logged in?
- How do I respect Concrete CMS permissions?
- How do I prevent the AI from exposing private pages?
- How do I search CMS content instead of dumping everything into a vector database?
- How do I keep AI responses synchronized with changing content?
Atlas separates responsibilities:
- Concrete CMS handles authentication, authorization, and content retrieval.
- Your AI agent focuses on understanding questions and generating answers.
This means you can use whatever AI framework you like—Mastra, LangGraph, OpenAI Agents SDK, or something completely different.
One last thing…
Some of you may remember that we previously released a Concrete CMS MCP Server.
Atlas Framework doesn’t replace it—they solve different problems.
-
Concrete CMS MCP Server enables AI to work with Concrete CMS, opening the door for AI-powered content management, automation, and editing.
-
Atlas Framework helps you quickly build AI assistants that can safely use your CMS content while respecting existing authentication and permissions.
They complement each other nicely.
Use Atlas to build AI assistants for your users.
Use MCP when you want AI to become a CMS user itself.
Hopefully this saves you a few weekends. 