A 0-100 composite score reflecting an MCP server's reliability based on runtime behavioral data (60%) and static metadata (40%). Scores above 70 indicate reliable servers. Updated continuously as new interaction data arrives.
Behavioral Baseline
Statistical norms (mean, p50, p95, min, max) for a category of MCP servers, computed from aggregated interaction data. Used to determine whether individual server performance is within normal range.
MCP Drift
Gradual degradation in an MCP server's reliability metrics over time, detectable by comparing current performance against historical trust scores and category baselines.
Trust Decay
The reduction in trust score confidence when an MCP server has not been observed (via probes or agent reports) for an extended period. Servers with no data in 30+ days have reduced score confidence.
Cross-Ecosystem Telemetry
Runtime behavioral data collected from multiple independent agents across different frameworks (LangChain, AutoGen, CrewAI, etc.) calling the same MCP server. This cross-ecosystem view is what distinguishes Observatory from single-platform metrics.
Compliance Attestation
Machine-readable audit trail of agent-MCP interactions, formatted for EU AI Act Article 12 (logging/traceability) and Singapore IMDA Agentic AI Governance Framework requirements.
Anomaly Detection
Real-time comparison of observed server behavior against historical performance and category baselines. Flags interactions that deviate beyond 2 standard deviations from expected norms.
Observatory Probe
Active monitoring request sent by Observatory to MCP servers on a scheduled basis (every 15 minutes). Probes test endpoint availability and response time independently of agent traffic.
Agent-Reported Interaction
A telemetry record submitted by an AI agent after calling an MCP server, via the Observatory SDK or direct API. Contains: server URL, success/failure, latency, tool name, and error details if applicable.
Trust Gate
A pre-flight check that verifies an MCP server's trust score meets a minimum threshold before an agent routes a call to it. Available in the dominion-observatory-langchain package as trust_gate().