Agent Standings

Methodology and probe policy

What we measure

Public remote MCP servers listed in the official MCP Registry (latest version, status active, an https remote without URL variables). Servers under domain-verified namespaces are added first; servers that someone looks up through check_tool are added on request (bounded). We don't rank our own server.

Probe policy

The request we send

POST <endpoint>
User-Agent: AgentStandingsProbe/0.1 (+https://agentstandings.com/probe)
Content-Type: application/json
Accept: application/json, text/event-stream

{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"AgentStandingsProbe","version":"0.1.0"}}}

Status definitions

Uptime = up / (up + down), where up = up + auth required + degraded and down = down + error. An incident opens at the first down or error check and closes at the next up check. Latency is the initialize round trip; p50 and p95 are per UTC day. Drift: we hash each tool's name, title, description, input and output schema and annotations (key order and list order don't count), and log a drift event whenever the set changes between checks.

Neutrality

Corrections log

No corrections yet.

Opt out

Stop all probes to a host. We stop within 24 hours (normally at once) and show the server as "not measured (opt-out)".