apoer-agent-guardrail
由 uchibeke 提供
Coding Agents & IDEs简介
Pre-action authorization for AI agents.
描述
该技能的功能
APort Agent Guardrail intercepts every tool call an AI agent makes and checks it against a policy before the tool runs. The policy engine is deterministic and fail-closed — if the guardrail errors or denies a call, the tool is blocked. Install once; the OpenClaw plugin enforces policy automatically after that.
The guardrail runs in before_tool_call, so the agent cannot skip or work around it, unlike post-hoc logging or advisory-only checks.
何时使用
- Blocking unauthorized shell commands before an agent executes them
- Preventing message sends to unapproved recipients
- Auditing MCP tool calls across automated workflows
- Requiring policy approval before PR creation or merge
- Stopping accidental data exports in production environments
示例工作流
User asks: An AI agent attempts to send a WhatsApp message to an external number
- 1
Agent prepares to call the messaging tool with target number and message content
- 2
APort OpenClaw plugin intercepts the call in before_tool_call
- 3
Guardrail script checks tool name 'messaging.message.send' and context JSON against the local or hosted passport policy
- 4
Policy engine returns ALLOW (exit 0) or DENY (exit 1)
- 5
If denied, the tool is blocked and the reason is written to decision.json
Tool either runs on ALLOW or is blocked, with the denial reason logged to decision.json
前置要求
- aport.io account and agent_id (optional, required only for hosted passport mode)
安装
clawhub install uchibeke/apoer-agent-guardrail运行 CLI 命令,或把这段提示词粘贴到 Claude Code、Cursor、Codex 或你在用的任意编程智能体里。