AgentGuard

Constitutional Governance Kernel for AI Agents

"Your AI agents are only as safe as the system governing them." "If your agent can break AgentGuard, it better be good at breaking Bitcoin."

---

Install

pip install agentguard-kernel

---

The Problem

You're building with AI agents. Your agent has access to your database, your payments, your users.

What happens when it does something it shouldn't?

Most teams have no answer. No pause button. No audit trail. No trust scoring. No approval gate. Just an agent with full access and a prayer.

AgentGuard changes that.

---

Quick Start

from agentguard import TrustEngine

engine = TrustEngine(db_path="trust.db")

# New agent starts TRUSTED
level = engine.get_trust_level("my-agent-001")
print(level)  # → TRUSTED

# Quick permission check
if engine.is_allowed("my-agent-001"):
    # proceed with action
    pass

# Agent does something bad
result = engine.apply_penalty("my-agent-001", "spam")
print(result)  # → {"new_points": 20, "new_level": "LIMITED"}

# Agent does something worse
result = engine.apply_penalty("my-agent-001", "abuse")
print(result)  # → {"new_points": 50, "new_level": "OBSERVED"}

# Agent recovers with good behaviour
engine.upgrade_trust("my-agent-001", reason="passed_review")

# Emergency: hard revoke
engine.revoke("my-agent-001", reason="critical_violation")

# Full audit trail
history = engine.get_history("my-agent-001")

# See all agents
agents = engine.list_agents()

---

Trust Levels

TRUSTED → LIMITED → OBSERVED → SUSPICIOUS → QUARANTINED → REVOKED

Level	Points	What it means
TRUSTED	0–19	Full access, normal operation
LIMITED	20–39	Restricted actions, monitored
OBSERVED	40–59	All actions logged, some blocked
SUSPICIOUS	60–79	Most actions require approval
QUARANTINED	80–99	Almost fully restricted
REVOKED	100	Completely blocked

---

Violation Penalties

Violation	Points
fake_verification	50
hate_speech	60
drug_listing	40
unauthorized_access	40
high_risk_activity	35
abuse	30
multiple_flags	25
spam	20

---

API Reference

TrustEngine(db_path="trust.db")

Initialize with a SQLite database path. Uses trust.db by default.

get_trust_level(user_id) → str

Returns the current trust level string.

get_trust_score(user_id) → dict

Returns {"level": ..., "points": ..., "appeal_status": ...}.

is_allowed(user_id) → bool

Returns False for QUARANTINED or REVOKED agents. Use this as your fast gate check.

apply_penalty(user_id, violation_type, details="") → dict

Applies a penalty. Returns {"new_points": ..., "new_level": ...}.

upgrade_trust(user_id, reason="") → dict

Rewards good behaviour, reduces points by 15.

revoke(user_id, reason="") → dict

Hard stop — immediately sets agent to REVOKED.

submit_appeal(user_id, reason, evidence) → dict

Agent submits an appeal for review.

get_history(user_id, limit=50) → list

Full audit trail for an agent.

list_agents() → list

All tracked agents and their trust state.

---

Architecture

┌─────────────────────────────────────────────────┐
│                   Your App                       │
├─────────────────────────────────────────────────┤
│         AgentGuard Kernel                        │
│    Trust Engine │ Risk Scoring │ Audit Log       │
├─────────────────────────────────────────────────┤
│              SQLite (Community)                  │
│         PostgreSQL (Enterprise)                  │
└─────────────────────────────────────────────────┘

Zero dependencies. Drop into any project in 5 minutes.

---

Enterprise Edition

The community kernel is free and open source.

Enterprise adds: REST API, JWT auth, Ed25519 capability tokens, human approval gate, real-time event stream, dashboard UI, global emergency stop, multi-tenant orgs, PostgreSQL + Redis, SSO/SAML, webhooks, SLA, and priority support.

📧 bkdk62309@gmail.com — response within 24 hours.

---

License

MIT — free to use, modify, and build on.

---

Author

Dheeraj Kumar Biswakarma 🐙 github.com/Mangomindai/agentguard