If you build agents, you've probably hit this: a site advertises MCP support, but when you actually wire it into Claude Code, Cursor, or your own client, the /mcp endpoint either doesn't exist or doesn't answer JSON-RPC correctly.
Q2 2026 data: of 5,578 sites advertising MCP support, only 10.3% (575) actually respond to a live tools/list request. Text-mention claims outnumber working endpoints roughly 4:1. Source: 8bitconcepts MCP ecosystem health report.
Here are three one-liners to verify any MCP server before integrating.
curl -sX POST https://nothumansearch.ai/mcp \
-H 'content-type: application/json' \
-H 'accept: application/json, text/event-stream' \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' \
| jq '.result.tools[] | {name, description}'Replace https://nothumansearch.ai/mcp with whatever /mcp URL you're checking. If you get a JSON body with a .result.tools array, the server is legit. If you get HTML, a 404, or "method not found", it's not a real MCP server regardless of what the site says.
If you don't want to deal with JSON-RPC headers, use the public verify endpoint:
curl -s "https://nothumansearch.ai/api/v1/verify?url=https://example.com/mcp"Returns the tool list and latency. Free, no auth.
MCP is only one of five signals that matter. llms.txt, OpenAPI, ai-plugin.json, and structured APIs count too. Score any site in one curl:
curl -s "https://nothumansearch.ai/api/v1/check?url=https://your-site.com" | jqIf you're picking MCP servers to wire into production, the "advertised vs actually works" gap is real. Text-mention-only "MCP-enabled" claims outnumber working endpoints 4:1. Verifying before integrating saves debug cycles downstream.
- Q2 2026 MCP ecosystem health report — methodology + raw CSV
- Not Human Search — live index of 1,929 agent-ready sites, ranked by agentic readiness score
- AI Dev Jobs — 8,405 AI/ML roles, 489 AI-first companies