SAFETY · OPEN SOURCE

AgentQ

AI Agent Observability, Safety, & Guardrail Control Plane

🤖
Your Agent
🛡️
AgentQ
21 Rules · 5 Classes
📊
Dashboard
🔔
Alerts
THE PROBLEM

AI agents in production have no safety net

Prompt injections, data exfiltration, goal drift, and scope creep all reach your users before you know they happened. Standard logging captures what agents did not whether it was safe.

🛡️
21

Guardrail Rules

Across 5 threat classes, evaluated on every span

🔌
0

SDK Changes Required

Drop-in OTel/MCP receiver — just point your endpoint

⚠️
5

Threat Classes

Injection, scope, exfiltration, behavioral, integrity

CAPABILITIES

Real-time observability for agent spans

🛡️

21 Guardrail Rules

Injection, scope creep, data exfiltration, behavioral anomalies, and integrity violations detected in real time on every ingested span.

🔌

Drop-in OTel Receiver

Set OTEL_EXPORTER_OTLP_ENDPOINT to point at AgentQ; MCP spans normalized automatically, no SDK changes required.

Tool Execution Interceptor

POST /api/intercept before any tool fires; surface violations before side effects occur.

🔬

Behavior Clustering

Composite embeddings (structural + semantic, all-MiniLM-L6-v2); LLM rubrics auto-generated at 10 traces per cluster.

🔔

Multi-Channel Alerts

Webhook, Slack, email; per-rule frequency caps and cooldown windows to prevent alert fatigue.

📊

Live Dashboard

SSE-driven trace feed, waterfall timeline, DAG graph, service graph, violation dashboard, and episode replay.