How chrismsmr-celcom/cerbere-ag works
Cerbère is an AI-agent runtime security and observability control plane. Its principal workflow wraps an agent’s model and tool calls in the SDK, evaluates security and policy checks, then reports spans to the Collector for storage, APIs, and dashboard views. The graph also preserves identity and audit capabilities and the documented MCP integration. Detection details and some API routes are documented or indexed but not fully sampled, so those links are limited to supported references.
Main entry points: policy.py SDK policy checks · sdk.py AgentGuard SDK · app.py Collector application · db.py Collector database
Components 26
Agent runtime
- AgentGuard SDK
agentguard/sdk.py - Runtime checks
agentguard/runtime.py - Taint tracking
agentguard/taint.py - Security models
agentguard/models.py
Security decisions
- SDK policy checks
agentguard/policy.py - Rules and patterns
agentguard/patterns.py - Input normalization
agentguard/normalizer.py - ML detection
agentguard_ml.py - LLM judge
judges.py - Unified decision engine
decision_engine.py - Risk scoring
risk_engine.py - Policy engine
policy/engine.py
Collector and observability
- Collector application
collector/app.py - Collector API
collector/api/__init__.py - Span ingestion
collector/api/spans.py - Collector database
collector/db.py - Dashboard views
collector/api/dashboard.py - Metrics API
collector/api/metrics.py
Identity and audit
- Identity management
collector/identity_routes.py - Authentication
collector/auth.py - Audit API
collector/audit_routes.py - Immutable audit log
audit.py
Agent integrations
- MCP server
mcp/mcp_server.py
Other components
- AI agent
- Collector service
- Dashboard user
Connections 30
- AI agent → AgentGuard SDK: invokes
README.md - AgentGuard SDK → Runtime checks: checks calls
agentguard/sdk.py - AgentGuard SDK → SDK policy checks: evaluates policy
agentguard/sdk.py - AgentGuard SDK → Taint tracking: tracks taint
agentguard/sdk.py - Runtime checks → Security models: uses models
agentguard/runtime.py - SDK policy checks → Security models: uses models
agentguard/policy.py - SDK policy checks → Rules and patterns: checks patterns
agentguard/policy.py - SDK policy checks → Input normalization: normalizes input
agentguard/policy.py - SDK policy checks → ML detection: classifies input
agentguard/policy.py - SDK policy checks → LLM judge: judges ambiguity
README.md - Unified decision engine → Risk scoring: scores risk
decision_engine.py - Policy engine → SDK policy checks: enforces capabilities
- AgentGuard SDK → Collector service: sends spans
agentguard/sdk.py - Collector service → Collector application: serves requests
- Collector application → Collector API: registers routes
collector/app.py - Collector API → Span ingestion: routes spans
collector/api/__init__.py - Span ingestion → Collector database: writes events
collector/api/spans.py - Dashboard views → Collector database: reads data
collector/api/dashboard.py - Metrics API → Collector database: reads metrics
collector/api/metrics.py - Dashboard user → Dashboard views: views activity
README.md - Collector application → Authentication: registers auth
collector/app.py - Collector application → Identity management: registers routes
collector/app.py - Identity management → Authentication: resolves identity
collector/identity_routes.py - Identity management → Collector database: stores identities
collector/identity_routes.py - Collector application → Audit API: registers routes
collector/app.py - Audit API → Authentication: requires auth
collector/audit_routes.py - Audit API → Immutable audit log: queries audit
collector/audit_routes.py - Audit API → Collector database: checks database
collector/audit_routes.py - AI agent → MCP server: uses MCP
README.md - MCP server → AgentGuard SDK: imports SDK
mcp/mcp_server.py