MDN's MCP Server Is a Reminder That AI Agents Are Only as Good as Their Sources

Mozilla has been quietly running an MDN MCP server as an experiment since June, giving AI coding agents and IDEs direct access to MDN's own documentation and browser compatibility data through the Model Context Protocol rather than whatever the model happened to absorb during training. It has not made much noise, but it addresses a problem every developer using an AI coding assistant has hit without necessarily naming it.

Models trained on data with a cutoff date do not know about a CSS property that shipped last month, or that a browser just changed its support status for an API it has carried for years. Ask a coding agent about scroll-axis-lock or the new camera element in Chrome 153 without an MCP connection to current documentation, and it will either invent a plausible sounding answer or quietly fall back on outdated information from its training data. Neither failure mode announces itself, which is what makes it dangerous.

We treat this as a general rule now, not just an MDN specific one: any AI tool we use for web development work needs a live connection to current documentation, not just a well trained model. A model is only ever as current as its training cutoff. A well maintained MCP server pointed at a source like MDN is current as of this afternoon. For anything touching fast moving web platform features, that difference is the whole reason to bother connecting the two at all.