MCP Server
Point any MCP client at the NachUI server and your agent can list the catalog, read the real source of a component and search the docs before it writes a line.
An assistant that guesses what a component looks like will invent one that does not exist. The MCP server closes that gap: it exposes the same registry the docs and the CLI use, so the agent reads the actual file before it writes anything.
Nothing is installed and nothing runs in the background. It is four read-only tools over the registry.
Setup
Add the server to your MCP client. For Claude Code, Cursor and most others, that is an entry in the client's
mcp.json:The package is fetched on first use and cached by
npx. Restart the client and the tools show up.Tools
| Tool | What it returns |
|---|---|
list_components | Every component in the registry, with the npm packages it needs. |
get_component | The source of one component, the same file you would copy by hand. |
search_docs | Semantic search across the documentation, in English or Spanish. |
get_doc | A full docs page as markdown, for when a snippet is not enough. |
A typical turn: the agent calls
list_components to see what exists, search_docs to find how a pattern is meant to be used, and get_component to pull the file it is about to adapt. You get the same copy and paste as always, done by the agent, from the real source.Slugs
Components are addressed by family and name, the same slugs the CLI takes:
ui/button, layout/stack, ai/prompt-input, hybrids/chat, icons/search. A bare name like button works when it is unique across families.Related
Found something to improve?
Notice a bug, typo, or missing detail on this page? Help us make the documentation better by opening a GitHub issue.