MCP server · an Upstash project · MIT licensed

Stop shipping
hallucinated APIs.

Context7 pulls up-to-date, version-specific documentation straight from the source and injects it into your AI's prompt — so Claude, Cursor and Codex answer with the API that exists today, not the one from a training crawl.

Try the index live demo

⏎

your-agent · prompt context

// waiting for a query…

resolve-library-id → query-docs → inject

Read the docs ↗ Explore the index ↓

0libraries indexed

0tokens indexed in this 290-library sample

0GitHub stars

0MCP clients supported

01 · the problem

Your AI's training data expired.

Models answer from a months-old snapshot of the internet. Libraries don't wait — they ship breaking changes weekly. The result: confident answers about APIs that no longer exist.

training cutoff · stale

Add rate limiting to my Express API with Upstash Redis.

✕ hallucinated — no package named express-rate-limit-redis exists on npm; the model blended express-rate-limit with a store API it invented, and the RedisStore constructor shown has never existed in any published version.

docs fetched 6 hours ago · v2.0.x

Add rate limiting to my Express API with Upstash Redis. use context7

✓ grounded — every symbol above was resolved from the current @upstash/ratelimit docs, fetched by Context7 at prompt time.

02 · how it works

One phrase. Fresh docs in the prompt.

Say use context7 (or set a rule once). Before your agent writes a line, Context7 resolves the exact library, pulls the current version's docs, and injects them into the context window.

1
Resolve. Your question is matched against 126,000+ indexed projects — fuzzy names, aliases and versions all land on an exact Context7 ID.
2
Fetch. query-docs pulls the current, version-matched snippets — parsed, scored and ranked for your specific question.
3
Inject. The docs enter the prompt before generation starts. The model reads the real API, then writes code that compiles.

03 · coverage

A constellation of everything you build with.

Every star below is a real project in the Context7 index — brightness is trust, size is community, drift is the index re-crawling. Search it, tap a star, steal its Context7 ID.

04 · freshness proof

Recency is a score, not a promise.

Every indexed project carries a trust score, a benchmark grade and a last-crawl timestamp. Here are the numbers the index reports right now — from the same snapshot powering the constellation above.

Benchmark = answer quality on real queries · Trust = source authority + crawl health · “updated” = last successful re-index.

05 · clients

Already wired into your editor.

Any MCP-capable client can reach the hosted server at https://mcp.context7.com/mcp. One CLI command configures the popular ones for you.

ClientOne-command setupRemote MCPLocal server

06 · install

Thirty seconds to current docs.

# requires Node.js 18+ — authenticates, generates an API key,
# and installs the right skill or MCP entry for your agent
$ npx ctx7 setup

# target a specific agent
$ npx ctx7 setup --claude
$ npx ctx7 setup --cursor

{
  "mcpServers": {
    "context7": {
      "type": "http",
      "url": "https://mcp.context7.com/mcp",
      "headers": { "Authorization": "Bearer YOUR_API_KEY" }
    }
  }
}

# no install at all — end any prompt with two words

Create a Next.js middleware that checks for a valid JWT
in cookies and redirects to /login. use context7

# or pin an exact library and skip resolution

Implement auth with Supabase. use library /supabase/supabase

07 · open source