Context7 MCP Server

Every developer using AI assistants knows the failure mode: confident code against an API that changed two majors ago, imports from packages that don't exist, config options the library never had. The model's training data froze; the ecosystem didn't. Context7 attacks exactly this — on request, it pulls current, version-specific documentation and working code examples for the library you're using and injects them into the model's context.

The workflow is almost invisible, which is its charm. Add "use context7" to a prompt (or let your client's rules invoke it automatically) and the assistant first resolves the library — Next.js 15, say, not Next.js-in-general — then fetches focused documentation for the topic at hand before answering. The difference shows immediately with fast-moving frameworks: App Router idioms instead of pages-directory fossils, current Tailwind syntax, today's Supabase client instead of 2023's.

Under the hood, Upstash maintains parsed, indexed documentation for tens of thousands of libraries, refreshed continuously. The two tools are appropriately boring: resolve-library-id finds the right library, get-library-docs fetches docs for it, optionally scoped to a topic like "routing" or "authentication".

It's free for personal use without an API key, though creating one raises rate limits. Available remotely (mcp.context7.com) or locally via npm. If you install only one documentation-related MCP server, this is the category winner — cheap to run, zero workflow disruption, and it fixes the single most common way AI coding assistance goes wrong.

What trips people up

• Resolution sometimes lands on a similarly named package. If you already know the ID — /vercel/next.js, /supabase/supabase — hand it over directly and skip the lookup step; version-pinned IDs work the same way.
• The topic argument matters more than people expect. "Routing" or "middleware" returns a tight, usable slice, while omitting it returns a broad overview that crowds your own code out of context.
• Typing "use context7" on every prompt gets old fast. Put the instruction in your Cursor rules or CLAUDE.md and it fires whenever the model reaches for library specifics.
• Private and internal packages aren't in the index. Public repos can be submitted on context7.com, and a context7.json at the repo root tunes how yours gets parsed.