mig-governance 0.1.0

The AI agent firewall. Graph-based content-level governance for any agent framework.

mig-governance

The AI agent firewall. Graph-based content-level governance for any automation.

AGT checks if your agent can use a tool.
MIG checks what your agent sends through that tool.

pip install mig-governance

---

Why MIG?

Your agent has permission to send emails. Great.
But it just attached the entire customer database.
AGT said ALLOW. MIG says DENY.

Your automation can trigger payments. Great.
But it just wired $50,000 to an unverified vendor.
Your workflow said proceed. MIG says APPROVAL REQUIRED.

No LLM in the decision loop. Deterministic. Graph-based. Fail-closed.

---

Quick Start

Three lines to govern any action:

from mig_governance import Governor

gov = Governor()
result = gov.validate("Send salary data to [e-mail]")

print(result.decision)     # DENY
print(result.risk_score)   # 90
print(result.policy_id)    # DEFAULT-DENY-001

Decorator — wrap any function:

from mig_governance import Governor, ActionDenied

gov = Governor()

@gov.guard
def send_email(to, subject, body):
    email_api.send(to, subject, body)

# Safe — executes normally
send_email("[e-mail]", "Meeting", "See you at 3pm")

# Dangerous — blocked before execution
try:
    send_email("[e-mail]", "Data", "SSN: 123-45-6789")
except ActionDenied as e:
    print(e)  # MIG DENIED: PII + external destination

Server mode — for any automation platform:

pip install mig-governance[server]
mig-governance serve

MIG is now running at http://localhost:8000/validate

Add this URL as an HTTP step in Zapier, Make.com, Relevance AI, n8n, or Power Automate. Your automation is governed.

# Test it
curl -X POST http://localhost:8000/validate \
  -H "Content-Type: application/json" \
  -d '{"text": "Read current sales report"}'
# → {"decision": "ALLOW", "risk_score": 10, ...}

curl -X POST http://localhost:8000/validate \
  -H "Content-Type: application/json" \
  -d '{"text": "Send salary data to [e-mail]"}'
# → {"decision": "DENY", "risk_score": 90, ...}

---

Three modes, one package

Mode	For	How
Library	Python developers	from mig_governance import Governor
Decorator	Framework developers	@gov.guard on any function
Server	Automation platforms	mig-governance serve → HTTP API

---

What MIG catches that others don't

Check	Microsoft AGT	MIG
Tool permissions	✅	—
Content/payload inspection	❌	✅
PII detection in actions	❌	✅
Risk scoring (0-100)	❌	✅
Operator approval workflow	❌	✅
Graph-based policy matching	❌	✅
Full audit trail	✅	✅
Deterministic decisions	✅	✅

MIG doesn't replace AGT. MIG is the layer on top — content inspection that permission checking can't provide.

---

Architecture

Action comes in
    ↓
┌─────────────────────────────────┐
│  8-Step Validation Pipeline     │
│                                 │
│  1. PII Detection               │
│  2. Action Classification       │
│  3. Payload Analysis            │
│  4. Semantic Matching (ChromaDB)│
│  5. Graph Policy Match (NetworkX│
│  6. Risk Scoring                │
│  7. Override Evaluation         │
│  8. Audit Logging               │
└─────────────────────────────────┘
    ↓
┌────────┬────────┬──────────┐
│ ALLOW  │  DENY  │ APPROVAL │
│ Safe   │ Blocked│ Needs    │
│ proceed│ stopped│ human OK │
└────────┴────────┴──────────┘

Powered by:

• NetworkX — graph-based policy matching (not flat rules)
• ChromaDB — semantic similarity (not just keywords)
• SQLite — full audit trail (every decision logged)

---

Works with any automation platform

Zapier → Add "Webhooks by Zapier" step → POST to http://localhost:8000/validate

Make.com → Add "HTTP Request" module → POST to http://localhost:8000/validate

Relevance AI → Add custom tool → HTTP POST to http://localhost:8000/validate

n8n → Add "HTTP Request" node → POST to http://localhost:8000/validate

Power Automate → Add "HTTP" action → POST to http://localhost:8000/validate

LangGraph → Use decorator:

from mig_governance.integrations.langgraph import mig_tool

@mig_tool(gov)
def my_tool(param):
    ...

---

Custom policies

Create your own policy pack:

{
    "name": "My Company Policies",
    "policies": [
        {
            "id": "MYCO-DENY-001",
            "description": "Block sending financial data externally",
            "action_type": "share_document",
            "direction": "DENY",
            "keywords": ["financial", "revenue", "salary", "budget"],
            "conditions": {"destination": "external"}
        },
        {
            "id": "MYCO-ALLOW-001",
            "description": "Allow reading any internal reports",
            "action_type": "read_data",
            "direction": "ALLOW",
            "keywords": ["read", "view", "report", "summary"]
        }
    ]
}

gov = Governor(policies="./my_policies.json")

---

Free vs Pro

Feature	Free (local)	Pro (hosted)
Graph engine	NetworkX	Neo4j
Embeddings	ChromaDB	sentence-transformers
PII detection	✅	✅
Payload analysis	✅	✅
Risk scoring	✅	✅
Audit trail	SQLite	Cloud DB
Drift detection	—	✅
Equipment profiles	—	✅
Semantic matching	Basic	Full Cypher
Dashboard	—	✅ Web UI
Price	Free	Contact us

# Free — runs locally
gov = Governor(policies="./policies.json")

# Pro — connects to hosted MIG engine
gov = Governor(api_key="gal_live_xxxxx")

---

Framework alignment

MIG architecture aligns with:

• NIST 800-207 — Policy Decision Point + Policy Enforcement Point
• OWASP Agentic Top 10 (2026) — mitigates ASI01, ASI02, ASI03, ASI06
• Anthropic Zero Trust for AI Agents — least agency, architecturally enforced
• IEC 62443 — zone-conduit enforcement for OT/ICS
• CISA Agentic AI Guidance — deterministic governance at execution boundary

---

Installation

# Core SDK
pip install mig-governance

# With server mode
pip install mig-governance[server]

# With LangGraph integration
pip install mig-governance[langgraph]

# Everything
pip install mig-governance[all]

---

API Reference

Governor(policies=None, api_key=None)

Main governance class.

• gov.validate(action, context=None) → Decision
• gov.guard → decorator for any function
• gov.get_audit(limit=50) → list of recent decisions
• gov.get_policies() → list of loaded policies
• gov.get_stats() → governance statistics

Decision

• .decision → "ALLOW", "DENY", or "APPROVAL"
• .risk_score → 0-100
• .policy_id → matched policy ID
• .checks → list of pipeline check results
• .flags → list of triggered flags
• .is_allowed / .is_denied / .needs_approval → bool

Server endpoints

Method	Endpoint	Description
POST	/validate	Validate an action
GET	/health	System status
GET	/policies	List policies
GET	/audit	Decision history
GET	/stats	Governance statistics
GET	/graph	Policy graph structure
GET	/docs	Interactive API docs

---

Built by

House of Galatine — Execution control for critical infrastructure and AI agents.

Built by a mechanical engineer who understands what happens when ungoverned commands reach controllers.

• Website: houseofgalatine.com
• Playground: houseofgalatine.com/playground
• Email: [e-mail]
• Patent: USPTO Provisional #63/821,489

---

License

Business Source License 1.1 (BSL)

Free to use for any purpose. Cannot be offered as a competing commercial governance-as-a-service product.

Converts to Apache 2.0 on June 1, 2029.

For commercial licensing: [e-mail]