aport-agent-guardrails: Authorize AI Agent Actions Before Execution

Overview

APort Agent Guardrails is a pre-action authorization tool for AI agents. It uses the Open Agent Passport (OAP) standard and runs checks through host hooks, plugins, and framework adapters before supported tools execute, addressing the risk that prompt instructions alone cannot prevent an agent from taking an unauthorized action.

It supports runtime integrations for Claude Code, Cursor, OpenClaw, LangChain, CrewAI, and other listed frameworks, as well as a GitHub Actions repository guard. Checks can use APort's API or a local evaluator. Coverage varies by host, and several command-hook integrations are explicitly beta.

Key Features

• Evaluates supported tool calls against passport identity and policy before execution.
• Provides API and local verification modes, with fail-closed behavior as the default.
• Offers runtime integrations across agent frameworks, including hooks, plugins, and Python adapters.
• Generates a GitHub Actions workflow for repository checks, with hosted enforcement available for protected branches.
• Records allow and deny decisions in audit files; API mode supports signed decisions.
• Includes enforce, warn, and observe rollout modes, with different enforcement guarantees.

Use Cases

• Agent developers can block risky shell, file, web, or other supported tool actions before a framework executes them.
• Teams using Claude Code, Cursor, OpenClaw, LangChain, or CrewAI can apply a shared passport and policy model across supported integrations.
• Repository maintainers can collect evidence about AI-assisted repository changes in GitHub Actions and enable hosted blocking when ready.
• Developers working offline or with local policy requirements can use the local evaluator instead of API verification, while accepting its narrower policy coverage.

Getting Started

For the GitHub repository guard:

npx @aporthq/aport-agent-guardrails github

For runtime integrations, run npx @aporthq/aport-agent-guardrails and choose a framework, or pass its name as an argument. The setup wizard requires Node.js 18 or newer. See the README and framework guides for configuration details.

License in Plain Words

No standard license was detected. Without a license, the default copyright rules apply: you may read the code, but not reuse it. Check the repository or ask the authors.

Considerations

• The project is active, with regular commits and releases, but measured contributor activity is concentrated in one person. That makes continuity and review capacity worth evaluating for security-sensitive adoption.
• Host integration maturity differs: Codex, Gemini CLI, and Goose are beta, n8n is setup-only, and opencode is gated pending plugin testing. Actual enforcement depends on each host's exposed hooks and mapped tools.
• Local evaluation is a narrower subset than API evaluation. API mode adds features such as signed decisions and hosted controls, so teams should compare the modes against their policy and operational needs.
• The hooks require jq, and audit logging requires python3; the README says missing required dependencies can deny tool calls.
• GitHub metadata reports the license as NOASSERTION, while the README and repository badge identify Apache 2.0. Confirm the repository's license file and terms before relying on that licensing claim.
• Cloud features, including signed receipts, a global kill switch, and team sync, are a separate APort service rather than part of the local open-source code.