Context7: pull current library docs into the agent, not the training cutoff
Context7 is a documentation tool for agents. It does not know your repo. It knows how to pull library docs that are newer than the model's memory.
What it does
The README's pitch is the usual failure: the model writes a Next.js API that was renamed, or a Supabase client call from an old major. Context7 resolves a library and returns current examples into the session.
You can run it as MCP (https://mcp.context7.com/mcp or a local server) or as a CLI-backed skill that shells out to ctx7. The skill path is aimed at agents that prefer commands over large MCP tool schemas.
It is a hosted product with a public repo. An API key is recommended for the hosted MCP. That is a vendor relationship, not just a file you copy.
Who it is for
Anyone whose agent keeps hallucinating framework APIs. Especially useful on fast-moving stacks: Next.js, Supabase, Stripe, Playwright.
Not useful as a substitute for AGENTS.md. Current docs will not tell the agent to plan first or to leave RLS enabled.
Our take
This is one of the few MCP servers that pays rent on a typical coding session. Wrong APIs waste more time than missing GitHub issue tools.
Limits: you are sending library questions to someone else's service. Docs can still be incomplete. The agent can ignore the fetched snippet. Context7 will not stop a bad architectural choice; it only refreshes the reference material.
We are not affiliated with Upstash. The repo is MIT; the hosted service has its own terms.
Compared with AgentKit Drawer
None of our packs fetch live docs. They freeze rules that docs will never mention: do not disable RLS to unblock a query, do not grant entitlements on /success. Use Context7 beside a rules file, not instead of one.
No pack CTA here on purpose. This is infrastructure. Buy a kit if you need the never-do list; add Context7 if your model is behind the framework.