MCP Just Hit 1 Billion Downloads. Is Your Docs Site Ready?

The MCP TypeScript and Python SDKs crossed 1 billion total downloads this week. Across Tier 1 SDKs combined, the protocol is approaching half a billion downloads per month. Those aren't just installs sitting idle — they represent AI agents, coding assistants, and developer tools that are actively making tool calls.

A significant fraction of those calls are directed at documentation. Developers using Claude, Cursor, GitHub Copilot, and similar tools ask questions like "how do I authenticate with this API?" and "what does this config option do?" constantly. The agents behind those tools need to answer from somewhere — and if your docs site isn't ready, they'll either hallucinate or tell the developer to look it up themselves.

What "1 billion downloads" means for docs sites

The MCP ecosystem didn't grow slowly. The 2026-07-28 spec release — which made the protocol stateless and removed the need for sticky sessions — accelerated adoption significantly. Google published an analysis describing it as the biggest architectural change to MCP since launch. The practical result: more MCP servers are deployable in more environments, and the developer tooling that consumes them is growing fast.

Every developer using an MCP-aware tool is a potential agent querying your docs. Not a web crawler building a search index — an agent asking a specific question and expecting a specific answer with a citation. Raw HTML doesn't serve that use case. A 1.6 MB homepage with 47 script tags is not a useful answer to "what's included in your Enterprise plan?".

The 2026-07-28 spec and what changed

The July spec introduced several changes relevant to docs sites specifically:

Stateless protocol core. MCP servers no longer need sticky sessions. This means a hosted MCP endpoint for your docs — like the ones AgentReady generates — can run on any stateless infrastructure: edge functions, serverless, standard load-balanced deployments. The barrier to hosting a persistent MCP endpoint for your docs just dropped significantly.

Deprecation policy. The spec now has a formal deprecation lifecycle. Features go through a deprecated → removed timeline with at least one version of notice. This matters for anyone building on MCP-based docs infrastructure: the protocol is stable enough to build on.

Extensions framework. The Tasks and MCP Apps extensions are shipping. MCP Apps in particular is relevant for docs: it defines how agents request interactive capabilities from a server, with explicit user confirmation for side-effectful actions. If you want your docs to eventually support "file a support ticket" or "run this code example," that's the path.

What readiness actually looks like

A docs site that's ready for MCP-level agent queries has four things:

1. An indexed content store. Your pages are crawled, chunked, and embedded. When an agent asks a question, there's a vector search that returns the most relevant chunks — not a keyword search, and not a training-data guess.

2. A cited answer endpoint. The indexed chunks feed a RAG pipeline that synthesises an answer and cites its sources. The agent gets a specific answer with a URL it can verify, not a list of links to search through.

3. An MCP tool that wraps the endpoint. So agents using Claude Desktop, Cursor, or any MCP client can call ask_site("yourdomain.com", "how do I...") directly, without a custom integration.

4. An llms.txt and agent.json. Machine-readable maps that tell agents what your site covers and how to query it, before they send a single question.

What unreadiness looks like (and costs)

A docs site without these things isn't invisible to agents — it just returns bad signal. Agents will either:

• Fall back to training data, which may be months out of date
• Return "I don't have current information" and tell the developer to check the docs manually — defeating the purpose
• Hallucinate a plausible-sounding but wrong answer, with no citation to correct it

All three outcomes represent a developer who got a worse experience because of your docs, not despite them. At MCP-scale — half a billion tool calls per month across the ecosystem — those bad experiences add up fast.

The indexing gap in the current directory

AgentReady's public directory already has most major developer docs indexed: Stripe, Vercel, Next.js, Supabase, Tailwind, React, OpenAI, Anthropic, GitHub, Cloudflare, Linear, Figma, Resend, Fly.io. Those sites are queryable now via MCP — not because those companies did anything, but because they're popular enough that users indexed them.

The gap is in the long tail: every SaaS product, every open-source project, every internal docs site that developers spend time reading. Those aren't going to be indexed organically. If you maintain one of them, you need to add it yourself — ideally before your users start hitting agents with questions and getting training-data guesses back.

How long it takes

Submit a URL. AgentReady crawls the site, chunks the content, generates embeddings, produces an llms.txt, and exposes an MCP endpoint. The whole process takes under 5 minutes for a typical docs site. No code changes to your site. No DNS configuration. No SDK to install.

At 1 billion MCP downloads and growing, the question isn't whether agents will query your docs. They already are. The question is whether they'll get a cited answer or a hallucinated one.