Documentation

v1.0.0

SDK reference, REST API specification, CLI guide, and architectural concepts for AgentGuard.

Base URL: http://127.0.0.1:8000

Installation

AgentGuard is designed to be lightweight. The core package has zero runtime dependencies beyond Python 3.11+. Install only what you need.

Core (zero deps)

pip install agentguard

With dashboard server

pip install agentguard[server]

With LLM providers

pip install agentguard[openai,gemini,anthropic]

All extras

pip install agentguard[all]

Initialize the Control Plane

Create a single `AgentGuard` instance and use `@guard.protect()` to wrap any agent entrypoint. The decorator intercepts every outbound tool call and context exchange.

main.py

from agentguard import AgentGuard

guard = AgentGuard(mode="strict")

@guard.protect(agent_id="orchestrator_v1", trust_level="HIGH")
def run_agent(task: str) -> str:
    # Your agent logic here
    return agent.run(task)

Enforcement Modes

Three built-in enforcement modes control how strictly the policy engine acts. All modes produce full evidence records.

modes.py

# STRICT  — Block all policy violations. Zero AI override.
guard = AgentGuard(mode="strict")

# MONITOR — Log and alert but do not block. Safe for staging.
guard = AgentGuard(mode="monitor")

# AUDIT   — Read-only. Record decisions without enforcement.
guard = AgentGuard(mode="audit")