Agent OS
Agent Governance Toolkit (AGT) is a framework-agnostic governance layer for autonomous AI agents, enforcing policies, identity, sandboxing, and audit logging in deterministic code before actions reach external systems. It solves the problem of prompt-level safety being unreliable by making disallowed actions structurally impossible. It is for developers deploying AI agents in production who need policy enforcement, auditability, and compliance.
✨ Key features
- Policy enforcement via YAML, OPA, or Cedar rules
- Identity management with SPIFFE, DID, or mTLS
- Tamper-evident audit logging of all decisions
- Sandboxing with four privilege rings
- Kill switch, SLO monitoring, and chaos testing
- Multi-language SDKs: Python, TypeScript, .NET, Rust, Go
🎯 Use cases
- Govern tool calls in AI agents to block destructive actions
- Enforce human approval for sensitive operations like sending emails
- Audit agent decisions for compliance with regulators
- Integrate governance into MCP servers or Claude Code plugins
- Run OWASP compliance checks and prompt injection audits in CI
📦 Installation
🧰 Requirements: Python 3.10+, Node.js 18+, .NET 8+, Go 1.25+, Rust 1.70+ depending on SDK; optional Azure credentials for Azure-integrated features.
pip install "agent-governance-toolkit[full]"
For TypeScript: npm install @microsoft/agent-governance-sdk
For .NET: dotnet add package Microsoft.AgentGovernance
For Rust: cargo add agent-governance
For Go: go get github.com/microsoft/agent-governance-toolkit/agent-governance-golang
For Claude Code: add plugin marketplace and install governance plugin.
🚀 Usage
from agentmesh.governance import govern
safe_tool = govern(my_tool, policy="policy.yaml")
# policy.yaml
# apiVersion: governance.toolkit/v1
# name: production-policy
# default_action: allow
# rules:
# - name: block-destructive
# condition: "action.type in ['drop', 'delete', 'truncate']"
# action: deny
# description: "Destructive operations require human approval"
safe_tool(action="drop", table="users")
# GovernanceDenied: Action denied by policy rule 'block-destructive'
⚠️ Good to know
Public Preview: production-quality but may have breaking changes before GA.
❓ FAQ
What is the difference between the base package and the [full] extra?
The base agent-governance-toolkit installs only the compliance CLI; the [full] extra includes the consolidated core distribution with all governance modules.
Can I use AGT with frameworks other than Python?
Yes, AGT provides SDKs for TypeScript, .NET, Rust, and Go, all implementing core governance features.
How does AGT enforce policies?
AGT intercepts tool calls in deterministic application code before the action reaches the wire, evaluating policies and raising GovernanceDenied when an action is not allowed.
Is there a CLI for compliance checks?
Yes, the agt CLI provides commands like agt doctor, agt verify, agt red-team scan, and agt lint-policy.
📊 Repository
🤖 Overview, features, install steps and FAQ were generated from the project's README on Sep 4, 2026. Always check the original source before running commands.