Context7
MCP server hosting curated, version-correct documentation for 1,856+ frameworks and tools. Drop into any MCP host (Cursor, Windsurf, Cline, Claude Code) and the agent queries Context7 for the actual docs of whatever framework it’s using — instead of hallucinating APIs from training data. The wiki’s strongest data point for “the right way to do RAG is to curate the corpus.”
- Coverage: 1,856+ frameworks/tools at recording time (Supabase, Next.js, React, LangGraph, Pydantic AI, MongoDB, MCP itself, and the long tail)
- Transport: MCP server (JSON config drop-in)
- Pricing: free at the time of the source
- Setup: copy the JSON config from the GitHub README → paste into the host’s MCP config → refresh
Why it solves the AI-coding hallucination problem
The biggest failure mode of AI coding agents is hallucinated framework APIs — the agent confidently calls a function that exists in a different version, or doesn’t exist at all. Context7 fixes this by giving the agent a tool to look up the actual docs of the actual version it’s targeting. The agent decides retrieval depth per query (typically 5K-20K tokens).
Context7 is curated, not crawled. That’s the structural difference vs general-web RAG: someone selected the docs that matter and chunked them with framework awareness. The wiki’s CAG entry argues for the opposite (dump the whole corpus into context), and RAG vs Wiki argues for structured links. All three are valid; Context7 wins for the coding-agent-needs-to-look-up-an-API use case specifically.
Position in the wiki’s RAG-skepticism / RAG-defense thread
| Pattern | Best for | Example |
|---|---|---|
| Curated MCP RAG (Context7) | Coding agents needing exact framework APIs | This page |
| CAG | Bounded datasets that fit the context window | API doc dumps |
| LLM Wiki | Personal knowledge bases | This wiki itself |
| General web RAG | Massive unstructured corpora | Dify |
Context7 is the counterpoint to the wiki’s RAG skepticism: it shows that RAG works well when someone curates the corpus. The hallucination-fixing thesis is the part the wiki cares about — and it’s the part the other patterns can’t replicate cheaply.
Sources
- This MCP Server Will 10x Your Productivity (Cole Medin, 2025-04-20)
See Also
- MCP
- MCP server template — sibling Cole video (build your own)
- Cursor / Cline / Claude Code — primary MCP hosts
- CAG — the alternative for bounded datasets
- RAG vs Wiki