One open-source framework to enforce safety, cost, and compliance controls on any AI agent — LangGraph, CrewAI, Bedrock, Copilot, Databricks, Azure. Not a guardrail. Not a tracer. A control plane.
You deploy an agent on Monday. By Friday it has called the refund tool 4,200 times, leaked an API key into Slack, and cost $38K — none of which is in your dashboard.
Everything regulated enterprises need to ship agents safely — in one open-source library. Drop it into any framework. Every tool call governed by default.
# Drop into any existing agent. No framework rewrite.from acp import governfrom langgraph.graph import StateGraph graph = StateGraph(...)agent = govern(graph.compile(), "procurement-agent", budget=1.0) # That's it. All 24 controls now active:# ✓ Tool call interception ✓ Budget enforcement (80/100/200)# ✓ Policy engine (30 rules) ✓ Content scanner (34 patterns)# ✓ Drift + bias detection ✓ Red team + approvals# ✓ BU isolation + consent ✓ Dashboard reporting
Warden wrapped our entire NL2SPARQL agent graph with one call. Every tool invocation now has budget checks, content scanning, and audit trails — without touching our graph code.
| Warden | Guardrails AI | NeMo Guardrails | LangSmith | Bedrock Guardrails | Lakera | |
|---|---|---|---|---|---|---|
| 24 integrated controls | ✓ | — | — | — | — | — |
| Runtime budget enforcement | ✓ | — | — | tracking only | — | — |
| Self-evolving agents | ✓ | — | — | — | — | — |
| MCP-native governance proxy | ✓ | — | — | — | — | — |
| 6 framework adapters | ✓ | generic | generic | LangChain only | Bedrock only | generic |
| Autonomy tiers with approvals | ✓ | — | — | — | — | — |
| Red teaming + CVSS scoring | ✓ | — | — | — | — | basic |
| Business-unit isolation | ✓ | — | — | — | — | — |
| EU AI Act · GDPR mapping | ✓ | — | — | — | — | partial |
| Open-source · self-host | ✓ | ✓ | ✓ | — | — | — |
Install the SDK. Wrap your agent. See every tool call in your dashboard within sixty seconds.