Svelte moves faster than any model's training data, and the usual result is confident code written against an older API. Two of these four tools fix that directly — current docs on request, and an autofixer the model can run against its own output until the suggestions stop.
The Svelte team's own MCP server. It hands the model the up-to-date documentation for the sections it asks for, and statically analyses the Svelte code the model writes so it can fix its own output before you see it.
- list-sections — the index of every documentation section available, so the model picks rather than guesses
- get-documentation — the full current text of the requested sections, pulled from svelte.dev/docs
- svelte-autofixer — static analysis of generated code returning issues and best-practice suggestions; it is built to be called in a loop until nothing is left to fix
- playground-link — an ephemeral playground URL carrying the generated code, for testing something that was never written to a file. The code lives in the URL and nowhere else
- A doc-section resource you include yourself: the URI form svelte://slug-of-the-docs.md adds one documentation page to the session without asking the model to fetch it
- A svelte-task prompt that instructs the model how to use the server properly
No account and no key, either way you run it. Locally it is the npm package `@sveltejs/mcp` over stdio, launched with npx; remotely it is the URL https://mcp.svelte.dev/mcp, added as an http server. The docs state that the Svelte team does not log, store or inspect code sent to the remote server.
One command — npx -y @sveltejs/mcp