Does Context7 Really Kill API Hallucinations? Field Notes
I have a reflex now: every time an agent confidently writes me a hook_form_alter that stopped being the right pattern two Drupal versions ago, I sigh and reach for docs. Context7, Upstash's documentation MCP server, promises to make that reflex obsolete by piping current, version-pinned docs straight into the model's context. So I pointed it at the two things that break agents fastest in my world: fast-moving PHP frameworks and Drupal contrib churn. Does it really kill API hallucinations? Mostly. Sometimes. It depends who wrote the docs it's serving.
What Context7 actually is
Strip the marketing and it's an MCP server (plus a ctx7 CLI/skill layer) that exposes two tools to your agent: resolve-library-id, which maps a name like "laravel" to a Context7 ID, and a docs fetcher that pulls version-specific snippets from that library's source. You trigger it by appending three words to a prompt: use context7. It's MIT-licensed on the client side, sits at roughly 59k GitHub stars, and the free tier works without an API key (a key just raises your rate limits before you start eating HTTP 429s). The parsing backend, notably, is private.
Wiring it into Claude Code or Cursor is genuinely a two-minute job. The classic local config:
{
"mcpServers": {
"context7": {
"command": "npx",
"args": ["-y", "@upstash/context7-mcp"]
}
}
}Drop that in ~/.cursor/mcp.json, or run npx ctx7 setup and let it target your client. There's also a hosted endpoint at https://mcp.context7.com/mcp if you'd rather not spawn a local process.
The pitch versus my reality
Here's the thing the landing page won't tell you: Context7 doesn't read the internet. It serves docs that were indexed and parsed into its own store, much of it community-contributed. That's the whole ballgame.
For a tool whose entire promise is "no more outdated docs," it quietly inherits the same staleness risk one layer down — in who curated the index, and when.
For Symfony and Laravel this is great. They're popular, well-indexed, and versioned cleanly, so asking for the 7.x way to wire a service and getting the 7.x answer actually happened for me. The version pinning is real and it's the best part.
Drupal is where the shine came off. My field notes:
- Core is decent, contrib is a lottery. Ask about a headline module and you get something; ask about a mid-tier contrib module mid-migration and coverage thins out fast.
- Churn outruns the index. The 10-to-11 shift to
#[Hook]attributes is exactly the kind of change where a stale snippet is worse than no snippet, because it reads authoritative. - The long tail is a gap. Niche packages simply aren't indexed yet, and the agent doesn't always tell you it came up empty.
The token budget question
Every "use context7" call injects a chunk of docs into your context window, and it is not free. On a tight task the retrieved block can crowd out the actual code I want the model reasoning about. My rule of thumb: invoke it deliberately for an unfamiliar or fast-moving API, not as an always-on reflex. Treating it like a permanent middleware tax is how you burn tokens on documentation you didn't need.
Verdict
- Popular, well-versioned frameworks: yes, it meaningfully cuts hallucinated APIs.
- Drupal contrib and the niche long tail: verify everything; treat it as a hint, not gospel.
- Token-wise: aim it, don't leave it running.
It doesn't kill hallucinations. It moves the trust problem from "the model's memory" to "someone's index," which is a better place for it to live — as long as you remember the index has an author and a date.