The State of MCP — June 2026

A census of the Model Context Protocol ecosystem, built from a deduplicated index of every MCP server we can find across the major registries. All figures are measured, not estimated. Snapshot date: 2026-06-04.

TL;DR — five things we found

1. There are 22,561 MCP servers once you deduplicate across the major registries. The real number is bigger than any single registry shows, because the registries overlap heavily and double-count.
2. It's a developer-tools ecosystem first. Code/dev tooling is ~46% of every categorized server — bigger than the next five categories combined.
3. A third of the ecosystem is barely described. ~36% of servers are un- or loosely categorized. Metadata hygiene is poor; discovery is hard.
4. Almost nothing has a reliability track record. We could find independent behavioral/uptime data for just 93 servers — 0.4%. For 99.6% of MCP servers, you are installing a black box.
5. Big infra has planted flags, but the long tail is individuals. Microsoft, AWS, GitHub, Cloudflare, Stripe, Notion, Redis, Pinecone, Qdrant, HashiCorp and Neo4j all ship MCP servers — yet most of the 22.5k come from solo developers.

How big is it, really?

Every registry undercounts. PulseMCP, Smithery, the official list and others each show a slice, and they overlap. Deduplicated into one index, we count 22,561 distinct MCP servers across 22 categories. That makes MCP, ~18 months in, already larger than many mature plugin ecosystems — but far messier.

What is everyone building?

Of the 14,497 servers that carry a real category (excluding "uncategorized" and the catch-all "other"):

Takeaway: MCP today is overwhelmingly about giving coding agents more tools. The "agent does my whole job" categories (finance, health, productivity) are still small.

The metadata problem

~36% of all servers (8,064 of 22,561) sit in "uncategorized" or "other." Most ship with thin or missing descriptions and no topic tags. For a protocol whose whole promise is discoverability by an agent, that's a real gap — and an opportunity for whoever cleans it up.

The reliability blind spot

Of 22,561 MCP servers, we have independent behavioral data — does it respond, how often, how fast — for 93. That's 0.4%.

There is no public, independent reliability record for the MCP ecosystem. When you wire an MCP server into an agent, you're trusting a black box with no uptime history, no latency baseline, and no failure record. The ecosystem has optimized for quantity of servers and ignored whether they work. This is the gap we're now measuring, server by server. Check any server's trust score →

Who's building

The most-starred MCP projects are a mix of the protocol's own reference servers and a few breakout independents (Upstash's Context7, Browser MCP). But the signal that matters is the infrastructure incumbents who've shipped official MCP servers: Microsoft (Playwright), AWS, GitHub, Cloudflare, Stripe, Notion, Redis, Pinecone, Qdrant, HashiCorp (Terraform), Neo4j, MotherDuck. When the platforms commit, the protocol sticks. Beneath them: thousands of individual developers. MCP's center of gravity is still the solo builder.

Method & honesty note

Counts come from a deduplicated index of public registry and GitHub data, refreshed continuously. We deliberately do not publish a growth-over-time chart in this edition: our first-seen timestamps reflect when we ingested each source, not when servers were created, so a clean organic-growth series isn't yet honest to show. We're tracking it forward from here. Behavioral coverage (93 servers) is what we have independently measured to date and is expanding.