State of MCP Runtime Trust 2026: Only 18 of 22,561 Servers Can Be Independently Verified

2026-06-08 · Dominion Observatory

We maintain a deduplicated index of 22,561 MCP servers. We tried to independently verify all of them at runtime. Not by scanning the source in a repo, but by actually reaching the running server to check it responds and behaves. Only 18 could be checked.

Here is what that says about how we trust the tools our agents call.

The numbers

The rest are GitHub repos, npm packages, or local stdio servers. Code you can read, but not a running service anyone can verify in production.

Static scans read the code. They never see the running server.

The popular way to vet an MCP server today is a static scan: read the source in the repo, look for known issues, give it a grade. That is useful, but it grades the code in a repository. It is not the server your agent connects to at call time, and the two can differ.

A server can pass a code review and then, in production, be slow, dead, swapped, or behave nothing like its description. The attacks the security community worries about most for agents, tool poisoning and rug pulls, happen at runtime, after a human approved the server. That is exactly where a static scan cannot see.

The gap

So we have an ecosystem where 99.9% of servers cannot be independently reached or tested in production, and the dominant trust signal is a one time read of code that is not even the running artifact. That is not a reliability record. It is a black box with a nice README.

If you run agents in production, the question is not did this code pass a scan. It is can I prove what this server did the last thousand times an agent called it. Today, for almost every MCP server, nobody can.

What we measure instead

We measure MCP servers by behavior, not by reading their code. Every server we can reach gets tested for whether it responds, how often, how fast, and whether it does what it claims, over time, with a did:web and Ed25519 signed, hash-chained record so the history cannot be quietly rewritten.

It is a small slice of the ecosystem today because the ecosystem is structurally hard to verify. That is the point. The gap is the story, and it is what we are closing server by server.

Full live data: State of MCP report. Check any server: trust score lookup. How runtime verification works: liveness and revocation.


Check any MCP server's trust score: dominionobservatory.com/check

Browse all 22,500+ servers: Server Directory

← Back to all articles