Pierre MCP
An MCP server that gives AI design tools a memory.
Why it exists
A design team was using Claude every day to execute different kinds of tasks, and every session started from zero. It did not know the team’s naming, or the spacing decisions already argued about.
So Pierre stops treating that as a prompting problem. The model was not the bottleneck, the context was. What was missing was a place for the team’s accumulated judgement to live.
How it works
Pierre is an MCP server that sits between the designer and the AI. You can easily install it from the repo here. From then on it searches the team’s conventions before answering, and records new decisions as they get made.
It has four tools and stops there: search, read, record, list. Recall had to be automatic, because no designer remembers to ask for their own conventions, so the search tool is described in a way that makes the AI reach for it unprompted. Capture had to be a single call, because a documentation step with any friction never happens.
Pure markdown files
Underneath it is a folder of Markdown. One file per decision, frontmatter for metadata, no database and nothing to host. That was deliberate: knowledge you cannot open in a text editor is knowledge you will eventually lose.
It also means a team shares its thinking through git. You pull to inherit everyone’s decisions and push to share your own, and conventions arrive as diffs you review like any other change. Every rule carries the reasoning behind it, which is the part that lets the AI apply it to a case nobody anticipated.