For a Swift package that changed last week, that distinction is the whole point — the archive on disk is current and a model's memory is not. Pointing it at your own package's generated documentation is the case where it earns the most, because there is no public source for an assistant to have learned from. Note the search caps: on a very large archive a broad query is sampling rather than exhausting it, so narrow by archive and symbol kind when precision matters.
An MCP server over DocC, the documentation format Xcode and Swift-DocC produce. It reads archives already on your machine, so an assistant answers from the documentation for the exact version you have built rather than from whatever was in its training data.
- The archives it can see, with their metadata, so you know what is actually indexed — `list_archives`
- Search across all of them by query, narrowable to one archive and filtered by symbol kind — class, struct, enum, protocol and the rest — `search_docc`
- One symbol's full record by its documentation id, including declaration with syntax highlighting, abstract, references and platform availability — `get_symbol`
- Articles and tutorials as their own lookup, since DocC archives carry prose alongside symbols — `get_article`
- Structural browsing: walk an archive's hierarchy by path instead of guessing at symbol ids — `browse_archive`
- Caching of archives and symbols for repeated access, with archives loaded on demand rather than all at once
No account and no key — everything is local. Node.js: clone, `npm install`, `npm run build`. At least one `--archive-path` argument must be given or the server exits with an error; the flag can be repeated to cover several directories, so Xcode's derived data and a couple of Swift package documentation folders can all be indexed at once. Archives come from `swift package generate-documentation` or Xcode's Build Documentation. The package is `docc-mcp` (1.0.0 in package.json). Two performance limits worth knowing: results are capped at 50 per query, and search reads at most 100 files per archive.