Context7

What it is

AI models learn from a snapshot of the internet, so their knowledge of any library is months or years old. When a library changes, your agent keeps writing the old way, or invents functions that never existed. Context7 fixes that by looking up the current documentation and handing the relevant bits to your agent before it writes code.

It's made by Upstash, a cloud database company. You can use it two ways. One is as an MCP server, a plug-in that gives your agent new tools. The other is as a skill plus a small command-line tool called ctx7, which avoids MCP entirely. Both talk to the same hosted service.

One thing to know up front: the GitHub repo is only the connector. The part that crawls, parses and ranks documentation runs on Upstash's servers and is closed source. The README says so plainly, after a user complained that this wasn't clear.

What actually happens when you use it

1. You ask for something. For example, "add email sign-up with Supabase." You can add "use context7" to the prompt, but the setup command installs a rule so the agent reaches for it on its own when a library is involved.
2. It finds the library. The agent asks Context7 which library you mean and gets back an ID like /supabase/supabase. If you already know the ID, put it in your prompt to skip this step.
3. It fetches the relevant docs. The agent sends your question. Context7's server picks the matching snippets and code examples and sends back only those.
4. The agent writes code using current examples instead of its memory. Mention a version ("Next.js 16") and it tries to match that version.

Since a January 2026 rebuild, the filtering happens on Upstash's side. Their own figures say this cut the docs sent to your agent by about 65% on average.

What people like

• It's set-and-forget. One Hacker News user runs it as the official Claude Code plugin, always on, and has seen no downsides.
• Fewer errors on library upgrades. In a Cursor thread, one user called it a game changer that cuts errors when upgrading or using an older version of a library.
• It's fast. The same thread's short summary: Context7 for speed, Ref for depth, Ref being a rival docs service.
• Docs, not raw source. When someone asked why not just let Claude read the library's code, replies said docs explain when and why to use things, and can flag bugs fixed in later versions.
• It can save tokens versus web search. Upstash's own 100-question test found Context7 about 35% cheaper per question than Claude Code's built-in web search. Upstash ran it and didn't measure answer accuracy, so treat it as a vendor claim.

What people run into

• The free tier shrank. In January 2026 Upstash cut the free limits sharply. A co-founder explained that every request now runs through a model, which costs more. Today's plans page lists 1,000 free calls a month, then 20 a day. Pro is $10 a month for 2,000 calls.
• Quality is mixed. The top reply in that thread said it was never good: lots of tokens, partly wrong examples. Another said it flooded their context with half-relevant material. Most of these complaints came before or right around the January rebuild.
• It's only as good as the docs it indexed. One user always checks what it ingested and when, and sends direct links when it matters. Another opened the output, found it was the public docs with the formatting stripped, and uninstalled it.
• Some people find they don't need it. A few say it slows code generation and burns tokens without improving results. Others save the docs locally once and point the agent at those.
• "Rate limited" when you shouldn't be. Some editors drop the API key, so you get counted as anonymous. One Cursor user hit the limit after four requests on a fresh key. Cursor's marketplace plugin once ignored a paid plan entirely until Upstash fixed it in October.

Before you install

• It's a hosted service. Each lookup sends Upstash a short description of what you're working on, so it can rank results. In 2025 an early version asked agents to pass along your exact prompt, which Upstash removed after pushback. Keep secrets out of prompts that might trigger a lookup.
• Docs are an attack route. Anyone can submit a library, and its docs go straight into your agent. In February 2026 researchers showed a library's custom rules could tell agents to read .env files and send them out. Upstash fixed it within days. In May, Upstash itself tested a notice inside doc results that asked agents to offer to run a setup command, then switched it off after complaints. Keep your agent asking before it runs commands.
• Sign in. Anonymous use works, but you'll hit limits faster. The npx ctx7 setup route gives you a free key automatically.
• It needs Node.js 22.13 or newer for the ctx7 command-line tool.

The short version

Context7 solves a real and common problem: agents writing code for last year's version of a library. It's free for most solo builders, takes one command, and is one of the few add-ons many people leave on all the time. It isn't magic. Results depend on how well a library's docs were indexed, power users often prefer local copies or paid rivals, and you're trusting Upstash's servers with your questions and the content they send back.