MCP, the Model Context Protocol, is an open protocol for connecting AI applications to tools and data. Without a design system server, an agent writing UI works from the files it happens to open. If your Button isn't among them, it may write a new one. With a server, the agent can ask before it writes.
What the agent gets
FISYCO runs one such server. It is remote, so there is nothing to install: the agent signs in through your browser and then acts as you, seeing only the projects you can see. It has fourteen tools, grouped by when the agent needs them:
All tools only read, except the one that starts a library sync. check_ui_code reads the fragment and doesn't store it. Without an indexed repository (a code index), the agent still gets your tokens and team rules, but not your components.
Rules, not a file
Team rules list components to skip, components to reach for first, and text that must never appear in code. They shape the server's answers instead of sitting in a file the agent may not read. The pull request check applies the same rules, so an agent and a reviewer get the same note for the same code.
Related
- Connect an agent through MCP: Claude Code, Cursor, VS Code with GitHub Copilot, Windsurf.
- Agent rules and the agents report
- Your AI agent does not know your design system