57% of public MCP servers break directory rules

01What the registry looks like

On 6 November 2026, the official MCP Registry listed 57,747 servers, covering 28,572 remote hosts over Streamable HTTP.

So 10% of listed hosts in the sample didn't work as MCP servers at all: 67 didn't answer, and 139 answered with something else.

02Which rules fail most

Among the 1,113 servers with readable tools:

The other 478 (43%) pass everything this check can read.

Most of the failures are missing labels, not bad tools: 517 of the 635 failing servers fail only on titles and hints. The tool may work well, but the AI app and the reviewer can't tell whether it changes anything.

03Servers that ask for sign-in

These tools can't be read without an account, so only the sign-in setup was checked. Among the 675:

Each of these can stop an AI app from starting sign-in at all. How MCP sign-in should work.

04How to fix the most common failures

Most servers need only a few lines on each tool:

{
  "name": "search_products",
  "title": "Search products",
  "description": "Searches the catalog by keyword and returns matching products.",
  "annotations": {
    "readOnlyHint": true,
    "destructiveHint": false,
    "openWorldHint": false
  }
}

Set openWorldHint to true when a tool reaches people or services outside your product, such as sending an email. Set all three explicitly: ChatGPT's rules ask for each one, and Claude's need a title and a read-only or destructive hint. More on tool design.

05How this was measured

• Sample. Hosts were taken in a fixed pseudo-random order from the registry's latest versions, with at most 25 under one registered domain, so a single platform can't dominate. Each host was checked once. The 2,000 hosts came from 1,719 domains.
• Requests. Only initialize, tools/list and public metadata reads, three hosts at a time. No tool was called and nothing signed in.
• Left out. OpenAI's domain verification file, because a server only needs it once it's submitted.
• Rules. The check's rules were read from each directory's published documents in October 2026. Each result in the tool links to its source.
• Sensitivity. Counting one server per domain gives a similar result: 528 of 957, or 55%.

06What this doesn't show

Failing a rule isn't the same as being rejected, and passing isn't approval. Directories also review privacy, policy and sign-in with a real account, and test every tool live.

Many of these servers are side projects or internal tools that aren't trying to get listed, so this describes the registry, not submissions to ChatGPT or Claude. Hints and descriptions are checked by rules, not read by a person, so a few results will be wrong in either direction.

The script is public, and server names aren't published here.

07Check your own server

• In a terminal: npx mcp-directory-check https://example.com/mcp
• In CI: the GitHub Action fails a build when a rule breaks.
• In a browser: the MCP directory check, with no install or account.