2026-05-27 · Dominion Observatory
When an AI agent connects to an MCP (Model Context Protocol) server, it's trusting that server to respond reliably, return accurate data, and not fail silently. But how do you know if a server is trustworthy before your agent connects?
A trust score is a number from 0 to 100 that measures an MCP server's real-world reliability based on four factors:
Is the server reachable when probed? A server that goes down regularly will score lower, even if it performs well when it's up. Uptime is tracked continuously.
How fast does the server respond? Both average latency and p95 (worst-case) latency matter. Top-scoring servers maintain under 50ms average response times. Anything above 500ms degrades agent performance significantly.
What percentage of calls return a valid response? Servers that return errors, timeouts, or malformed responses on a significant portion of calls score lower. The best servers maintain 98%+ success rates.
Does the server's behavior drift over time? A server that was fast last week but slow today, or one that suddenly starts returning different response formats, gets flagged for inconsistency.
The trust score combines a static score (30%) based on registry metadata, documentation completeness, and GitHub activity, with a runtime score (70%) based on actual call performance. This means a well-documented server with poor runtime performance will still score low.
Your AI agent is only as reliable as the weakest MCP server it connects to. With an ecosystem average of 64.3 out of 100, most servers are mediocre. Checking trust scores before connecting is the simplest way to improve your agent's reliability.
Check any server's trust score: dominionobservatory.com/check
Check any MCP server's trust score: dominionobservatory.com/check
Browse all 22,500+ servers: Server Directory