Experiments with MCP in Wagtail
Prototyping with agents interacting directly with the CMS
Following our work to prototype a new CLI for Wagtail, it only felt right to experiment with other ways to give agents access to the CMS via our new API and open up opportunities for automation. And voilà, a new experimental MCP package!
Pretty cool, right? This is wagtail-mcp, vibe-coded with Kimi K3 and the Pi coding agent back in August, now stable enough for a demo.
50+ tools and an agent interface
The server comes with 50+ tools, same as the API. To demo this more easily, we also added an agent interface with the same tools directly in the CMS, built with the AG-UI protocol and CopilotKit. Here’s an example, asking the agent in natural language, how many redirects we have on the site in total:
We set this up with one of our preferred AI providers, Neuralwatt, so we know exactly how much energy this uses per query: less than 1 Wh - very little.
Beyond prototyping
Giving agents access to the CMS is very promising but also very dangerous, even with Wagtail’s strong permissions model. In a sense, adding MCP with authentication and permissions matching what we have for users was the easy part, and the hard work left to do is to decide how to make this available with stronger guarantees of safety. Whether we make it stricter for agents, or rework content operations in other ways.
In the meantime, if you want to experiment, do give a go to the new package, and let us know what you think!