Every MCP registry shows you a different slice, and they all double count. So I merged them. PulseMCP, Smithery, the official list, GitHub. Then deduplicated into one index. Here's the honest state of the MCP ecosystem about 18 months in. Everything here is measured, not guessed.
22,561 servers, for real
Once you dedupe across registries the real number is way bigger than any single one shows. 22,561 distinct servers across 22 categories. That already makes MCP bigger than a lot of mature plugin ecosystems. Just way messier.
it's a dev tools ecosystem, full stop
Code and dev tooling is 46% of every categorized server. Bigger than the next five categories combined. Right now MCP is mostly about handing coding agents more tools. The "agent runs my whole life" stuff (finance, health, productivity) is still tiny.
a third of it is basically a black hole
About 36% of servers are uncategorized or barely categorized. Thin descriptions, no tags. For a protocol whose whole pitch is "an agent can discover and use this," the metadata hygiene is rough.
almost nothing has a reliability track record
This is the one that should scare you if you're shipping agents to prod. I could find independent behavioral or uptime data for 93 servers. Out of 22,561. That's 0.4%. For the other 99.6% you're wiring a black box into your agent. No uptime history, no latency baseline, nothing.
the platforms showed up, but it's still mostly solo devs
Microsoft, AWS, GitHub, Cloudflare, Stripe, Notion, Redis, Pinecone, Qdrant, HashiCorp, Neo4j, MotherDuck all ship official MCP servers now. But the center of gravity is still thousands of individual devs.
what I didn't publish, on purpose
No growth over time chart. My first seen timestamps are when I ingested a source, not when a server was actually created. So a clean growth curve isn't honest yet. I'd rather show you nothing than fake a line going up and to the right.
Full report with the category breakdown and method notes is here: https://dominionobservatory.com/atlas/report
Top comments (0)