APort Agent Guardrail

在每次工具调用执行前进行确定性授权校验，覆盖 shell 命令、消息发送、git 操作、MCP 工具与数据导出等工具名映射；通过 OpenClaw 插件在工具调用前钩子自动强制执行，出错时默认拦截，并把决策记录为本地 JSON 或经由 APort API 的签名回执。

uchibeke

热度 27.81

最近更新 2026-05-17 12:12:49

分类：安全依赖：代码执行 / 读写文件 / 第三方 API应用场景：工作流 / 自动化脚本

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add apoer-agent-guardrail

Metadata

name: aport-guardrail
description: Pre-action authorization for AI agents. Verifies permissions before every tool runs (shell, messaging, git, MCP, data export). Works with OpenClaw, IronClaw, PicoClaw. APort policy engine allows or denies each tool call deterministically; agent cannot skip it.
homepage: https://aport.io
metadata: {"openclaw":{"requires":{"bins":["jq"]}}}

APort Agent Guardrail

Pre-action authorization for AI agents: every tool call is checked before it runs. Works with OpenClaw, IronClaw, PicoClaw, and compatible frameworks. Run the installer once; the OpenClaw plugin then enforces policy on every tool call automatically. You do not run the guardrail script yourself.

Requires: Node 18+, jq. Install with npx @aporthq/agent-guardrails or ./bin/openclaw from the repo.

Installation

# Recommended (no clone needed)
npx @aporthq/agent-guardrails

# Hosted passport: skip the wizard by passing agent_id from aport.io
npx @aporthq/agent-guardrails <agent_id>

Get a Hosted Passport agent_id at aport.io after creating a passport there. OPTIONAL

From the repo (clone first): github.com/aporthq/aport-agent-guardrails — then run ./bin/openclaw or ./bin/openclaw <agent_id> from the repo root. Full guides: QuickStart: OpenClaw Plugin · Hosted passport setup.

You can preview your local passport at ~/.openclaw/aport/passport.json (or <config-dir>/aport/passport.json if you chose a different config dir; legacy installs may use <config-dir>/passport.json).

The installer is interactive: it sets your config dir, passport (local or hosted), installs the APort OpenClaw plugin, writes config, and installs wrappers. After it finishes, nothing else is required—start OpenClaw (or use the running gateway); the plugin enforces before every tool call.

Wrappers (default config dir ~/.openclaw): ~/.openclaw/.skills/aport-guardrail.sh (local), ~/.openclaw/.skills/aport-guardrail-api.sh (API/hosted). The plugin uses these; you don’t call them unless testing.

Usage

Normal use: Run the installer once. After that, nothing to run manually—the plugin enforces before each tool call automatically.

Optional — direct script calls (e.g. testing or other automations):

~/.openclaw/.skills/aport-guardrail.sh system.command.execute '{"command":"ls"}'
~/.openclaw/.skills/aport-guardrail.sh messaging.message.send '{"channel":"whatsapp","to":"+15551234567"}'

Exit 0 = ALLOW (tool may proceed)
Exit 1 = DENY (see <config-dir>/aport/decision.json or <config-dir>/decision.json for reason codes)

For API mode / hosted passports:

APORT_API_URL=https://api.aport.io ~/.openclaw/.skills/aport-guardrail-api.sh system.command.execute '{"command":"ls"}'

Tool name mapping

When you're about to…	Use tool_name
Run shell commands	`system.command.execute`
Send WhatsApp/email/etc.	`messaging.message.send`
Create/merge PRs	`git.create_pr`, `git.merge`
Call MCP tools	`mcp.tool.execute`
Export data / files	`data.export`

Context must be valid JSON, e.g. '{"command":"ls"}' or '{"channel":"whatsapp","to":"+1..."}'.

Why this skill?

Deterministic – runs in before_tool_call; the agent cannot skip it.
Structured policy – backed by Open Agent Passport (OAP) v1.0 and policy packs.
Fail-closed – if the guardrail errors, the tool is blocked.
Audit-ready – decisions are logged (local JSON or APort API for signed receipts).

Pair it with other threat-detection tooling if needed; enforce policy through this guardrail so unsafe actions never run.

Docs

This repo: QuickStart: OpenClaw Plugin · Hosted passport · Tool / policy mapping

OpenClaw: CLI: skills · Skills · Skills config · ClawHub