Disclosure: I run MCP Verified, a weekly test of servers in the official Model Context Protocol registry. This post is the data. The newsletter link is at the end.
What we test
Each week we take a sample of 750 entries from the official MCP registry (300 npm, 150 PyPI, 200 remote streamable-HTTP, 100 remote SSE). We install or connect to each one in a clean sandbox, run the MCP initialize handshake, and call tools/list. Where possible we also make one harmless tool call. We supply no API keys.
Results (run of Oct 19, 2026)
| Type | Tested | Working | Median handshake | Median tools |
|---|---|---|---|---|
| npm | 300 | 218 (73%) | 563 ms | 14 |
| PyPI | 150 | 73 (49%) | 1,581 ms | 16 |
| HTTP | 200 | 130 (65%) | 436 ms | 8.5 |
| SSE | 100 | 10 (10%) | 1,152 ms | 9 |
| All | 750 | 431 (57%) |
Failure reasons (319 servers)
| Reason | Count |
|---|---|
| Needs config/env vars (untested) | 92 |
| Remote endpoint dead | 88 |
| Needs credentials (untested) | 54 |
| Crashed on start | 47 |
| Install failed | 12 |
| Handshake failed | 10 |
| tools/list failed after handshake | 6 |
| URL isn't an MCP server | 6 |
| Timeout (45 s) | 4 |
Practical checklist for server authors
- Don't crash when an env var is missing. Finish the handshake and return a clear error from the tool call.
- If you publish a remote endpoint, monitor it. 69 of 100 SSE entries were dead.
- Pin dependencies. Install failures were few (12), but each one is a total loss for users.
- Keep startup fast. Clients often time out. The PyPI median handshake was about 3× npm's.
Trend
Working count in our runs: 422 (Oct 13), then 427, then 431 (both Oct 19).
Caveats
This is a sample, not the whole registry. "Needs config" doesn't mean broken. Results are from one sandbox environment.
Weekly results (free newsletter): https://buttondown.com/mcpverified · Site: https://mcpverified.netlify.app
Top comments (0)