State of MCP servers · November 2026
We audited 12,596 MCP servers.
Most make agents guess.
We ran the same public check on one server from every domain in the official MCP registry: what an AI agent sees when it connects, and whether that is enough to use the server well. The median score is 76 out of 100.
Scores from 10 November 2026. Not listed, or changed since? Audit your server now: a few seconds, no signup.
01The numbers
40,629 servers listed.
12,596 scored.
3,582 servers score under 70. Only 48 reach 90. The other 9,004 domains have a server behind a login, which a public check can’t score, and 1,847 did not answer as an MCP server.
02What agents trip on
The same six gaps,
over and over.
Share of the 12,596 scored servers with each gap. None of them returns an error: the agent just guesses, and the guess shows up in your logs as a normal call.
Parameters with no description
The agent sees a name like time_window or filterKey and has to guess the format. Fix: one sentence per parameter, with the format and an example value.
Tool descriptions that don’t say when to use the tool
Descriptions that repeat the tool name or lean on words like “handles” and “manages”. With two similar tools, the model picks one at random. Fix: say what the tool returns and when to pick it over its neighbour.
Tools that don’t say if they read or write
Without readOnlyHint and destructiveHint, a client can’t tell a search from a delete, so it either asks the user every time or runs everything. Fix: set the hints on every tool.
No instructions when the agent connects
The server answers initialize with no instructions, so the agent starts with a list of tools and no idea which to call first. Fix: three or four sentences on what the server is for and where to start.
Fixed lists typed as free text
A parameter such as sort or lang accepts any string, but only a few values work. The agent invents one. Fix: declare the allowed values as an enum.
More than 40 tools
Past about 40 tools, models pick the wrong one more often and the list eats their context. Fix: group tools, or expose a search tool that returns the right one.
No way for the agent to say it got stuck
This one is ours, so read it with that in mind: the audit also checks for a feedback tool from the open Agent Feedback Protocol, and it counts for 15 of the 100 points. Fewer than 100 servers have one. It is why a server with no other gap tops out at 85.
2,350 servers (19%) have none of the first six gaps.
03Method
What we checked,
and what we didn’t.
How it was done
- Source: every active server with a remote address in the official MCP registry, read on 10 November 2026.
- One server per domain. Servers on shared hosting and domains publishing more than 20 servers were left out.
- About 20 public requests per server:
initialize,tools/list, and a read of each tool’s description, parameters and hints. No tool was called. - The audit is open source: @backloop/audit. Run it yourself and compare.
Limits
- It reads what a server declares. It does not test whether the tools work.
- Servers behind a login (9,004 domains) are not scored.
- A score is a snapshot. A server fixed since 10 November still shows its old score in the lookup; the live audit is current.
- The registry includes many small and experimental servers. This is a picture of what is published, not of the most used servers.
Your server
Not in the list?
Audit yours now.
Paste your MCP server or API address, new or already listed. You get the score, each gap and its fix in a few seconds. Free, no signup.
About 20 public requests. No tool is called and nothing is stored.