NARNA

Governance Infrastructure for Agentic AI

Govern Once. Run Anywhere.

NARNA makes Agentic AI governable — portable identity, policy, evidence, and trust across LangGraph, CrewAI, OpenAI, Anthropic, MCP, and OTel.

It does not replace those runtimes. It does not train models. It sits in the infrastructure layer and enforces / audits / proves what fleets may do.

NARNA governs AI. Others execute it.

Not another “agent passport” clone. See docs/DIFFERENTIATION.md — wedge = UGS + Packages + GU, not identity-only protocols.

Install

pip install -e .   # or: pip install narna  (when published)

from narna import wrap

# Enforce before host side-effects (default). Use mode="observe" to migrate.
agent = wrap(my_langgraph_app, vap=True, mode="enforce")
agent.run("quarterly summary")

30 seconds — package + runtime

from narna import Agent, ConstitutionRuntime

rt = ConstitutionRuntime()
rt.load(provider="eu-ai-act")  # Governance Package

agent = Agent(vap=True)
agent.run()

Why teams integrate NARNA

1. Compliance packages — load once, enforce across frameworks
2. Enforce-before adapters — deny tool/LLM calls before side effects
3. UGS Passport + Registry — public verify at narna.org
4. Cloud GU — Runtime free; Trust (Registry / Passport / Packages) is the product

Links

Compatibility

OpenAI · Anthropic · Google · MCP · OpenTelemetry · LangGraph · CrewAI · Docker · Kubernetes

License

MIT