Risk engine
Scores actions from context, sensitivity and impact.
Issue 01 MCP governance protocol
When AI
acts,
who watches?
A governance gateway that evaluates MCP actions, asks humans for approval when risk demands it, detects approval fatigue, and can hit the emergency stop.
InterceptAssessAuthorize
Human
in the
loopAuthority core
in the
loopAuthority core
Risk scan
Integrity
Override
Execute
Threat file 001
Too many
yes buttons.
Approval fatigue
Continuous approval requests can condition a reviewer to approve without meaningful scrutiny.
Human gate
Critical actions wait for an explicit human decision.
Circuit breaker
Suspicious request bursts can pause automated execution.
Chain of custody Zero blind spots
Every action
leaves a trail.
01MCP request
02Risk scan
03Human approval
04Hash / integrity
05Execute / block
System arsenal Six controls
Built to
intervene.
One gateway. Six coordinated safeguards between autonomous intent and real-world impact.
Risk engine
Scores each action from context, sensitivity, scope, and potential impact.
Approval manager
Suspends sensitive actions until an accountable human makes the call.
Fatigue detector
Flags rapid approval patterns before reflex replaces meaningful review.
Circuit breaker
Stops automated execution when request velocity becomes suspicious.
Hash manager
Verifies that the reviewed payload is exactly the payload that executes.
Audit logger
Preserves a legible trail of scores, decisions, checks, and outcomes.
Control is a choice