Imported from modbender/skill-library-mcp (data/aport-agent-guardrail/SKILL.md). Install upstream with npx skills add modbender/skill-library-mcp --skill aport-agent-guardrail. Copyright stays with the author.

APort Agent Guardrail

Skill identifier: aport-agent-guardrail · Category: Security / Infrastructure

🛡️ What This Skill Does

Pre-action authorization for AI agents. Every tool call is evaluated against a passport (identity + capabilities + limits) and policy before it executes. If denied, the tool never runs.

Key features:

• ✅ Deterministic enforcement – Runs in before_tool_call hook; agent cannot bypass
• ✅ Blocks malicious actions – Unauthorized commands, data exfiltration, API abuse prevented
• ✅ Structured policies – Based on Open Agent Passport (OAP) v1.0
• ✅ Fail-closed by default – Errors deny tool execution (security over availability)
• ✅ Audit trail – Every decision logged with tamper-evident hashes
• ✅ Framework-agnostic – OpenClaw, IronClaw, PicoClaw, and compatible runtimes

How it protects you:

• Prompt injection → Agent cannot bypass hook-based enforcement
• Malicious skills → All tool calls checked regardless of source
• Unauthorized commands → Allowlist + 40+ blocked patterns (rm -rf, sudo, etc.)
• Data exfiltration → File access, messaging, web requests controlled
• Resource exhaustion → Rate limits, size caps enforced

Install once, protected forever. The plugin runs automatically on every tool call.

⚡ Quick Start

# Install APort guardrails (one-time setup)
npx @aporthq/aport-agent-guardrails

# Follow wizard to create passport and configure policies
# Plugin auto-registers with OpenClaw

# Now your agent is protected
# All tool calls checked before execution

With hosted passport (optional):

# Get agent_id from aport.io
npx @aporthq/aport-agent-guardrails <agent_id>

Requirements: Node 18+, jq

📦 Installation

Option 1: npm (recommended)

npx @aporthq/aport-agent-guardrails

The wizard will:

1. Create or load passport (local file or hosted from aport.io)
2. Configure capabilities and limits
3. Install OpenClaw plugin automatically
4. Set up wrapper scripts

After install: Plugin enforces before every tool call. No further action needed.

Option 2: With hosted passport

npx @aporthq/aport-agent-guardrails <agent_id>

Get agent_id at aport.io for:

• Cryptographically signed decisions
• Global suspend (<200ms across all systems)
• Centralized audit and compliance dashboards
• Team collaboration

Option 3: From source

git clone https://github.com/aporthq/aport-agent-guardrails
cd aport-agent-guardrails
./bin/openclaw

Guides:

• QuickStart: OpenClaw Plugin
• Hosted passport setup

🚀 Usage

Automatic enforcement (default)

After installation, the plugin runs automatically:

# Your agent uses tools normally
agent> run git status
# ✅ APort: passport checked → policy evaluated → ALLOW → tool executes

agent> run rm -rf /
# ❌ APort: passport checked → blocked pattern detected → DENY → tool blocked

You do nothing. The plugin enforces on every tool call in the background.

Testing the guardrail (optional)

Direct script calls for testing or automation:

# Test allowed command
~/.openclaw/.skills/aport-guardrail.sh system.command.execute '{"command":"ls"}'
# Exit 0 = ALLOW

# Test blocked command
~/.openclaw/.skills/aport-guardrail.sh system.command.execute '{"command":"rm -rf /"}'
# Exit 1 = DENY

# Test messaging
~/.openclaw/.skills/aport-guardrail.sh messaging.message.send '{"channel":"whatsapp","to":"+15551234567"}'

Exit codes:

• 0 = ALLOW (tool may proceed)
• 1 = DENY (reason in decision.json)

Decision logs:

• Latest: ~/.openclaw/aport/decision.json
• Audit trail: ~/.openclaw/aport/audit.log
• API mode: Signed receipts via APort API

🔍 How It Works

Pre-Action Authorization Flow

User Request → Agent Decision → APort Check → [ALLOW/DENY] → Tool Execution
                                     ↑
                              Policy + Passport

1. User makes request (e.g., "Deploy to production")
2. Agent decides to use tool (e.g., exec.run with git push)
3. OpenClaw fires hook (before_tool_call)
4. APort evaluates:
   • Load passport (identity, capabilities, limits)
   • Map tool → policy (exec → system.command.execute.v1)
   • Check allowlist, blocked patterns, rate limits
5. Decision: ALLOW or DENY
6. Audit: Log decision with timestamp, policy, reason codes

Agent cannot bypass. Hook is registered by OpenClaw, not controlled by prompts.

What Gets Installed

Plugin registration:

• OpenClaw plugin added to config (enforces before_tool_call)
• TypeScript/JavaScript plugin loaded on OpenClaw start

Files created (under ~/.openclaw/):

• config.yaml or openclaw.json – Plugin configuration
• .skills/aport-guardrail*.sh – Wrapper scripts for local/API evaluation
• aport/passport.json – Your agent passport (local mode only)
• aport/decision.json – Latest decision (runtime)
• aport/audit.log – Audit trail (runtime)

Total disk usage: ~100KB scripts + your passport/decisions

Review code: GitHub repository

🌐 Network and Privacy

Local Mode (Default)

Zero network calls:

• ✅ All evaluation on your machine
• ✅ Passport stored locally
• ✅ Decisions stay local
• ✅ Full privacy
• ✅ Works offline

Perfect for: Development, personal use, air-gapped environments

API Mode (Optional)

Network usage:

• Tool name + context → APort API for policy evaluation
• Hosted passport fetched from registry (if using agent_id)
• Signed decisions returned (Ed25519 cryptographic signatures)

Benefits:

• ✅ Cryptographically signed decisions
• ✅ Court-admissible audit trail
• ✅ Global suspend across all systems
• ✅ Centralized compliance dashboards
• ✅ No local passport tampering possible

API endpoint: https://api.aport.io (or custom via APORT_API_URL)

Data sent:

• Tool name (e.g., "system.command.execute")
• Context (e.g., {"command": "ls"})
• Passport (if local) or agent_id (if hosted)

Data NOT sent:

• File contents
• Environment variables
• API keys or credentials
• Unrelated system information

To verify: Use local mode (no network) or inspect open-source code.

⚙️ Environment Variables

Local mode: No environment variables needed. Passport read from ~/.openclaw/aport/passport.json.

Hosted mode: Pass agent_id to installer or set APORT_AGENT_ID.

🔧 Tool Name Mapping

Context format: Valid JSON, e.g., '{"command":"ls"}' or '{"channel":"whatsapp","to":"+1..."}'

📋 Out-of-the-Box Protections

Shell commands (system.command.execute.v1):

• Allowlist enforcement (only specified commands run)
• 40+ blocked patterns: rm -rf, sudo, chmod 777, dd if=, mkfs, etc.
• Interpreter bypasses blocked: python -c, node -e, base64 encoding
• Command injection patterns detected

Messaging (messaging.message.send.v1):

• Rate limits (msgs_per_min, msgs_per_day)
• Recipient allowlist
• Channel restrictions

File access (data.file.read/write.v1):

• Path restrictions (block /etc, /bin, system directories)
• Prevent .env, SSH key theft

Web requests (web.fetch/browser.v1):

• Domain allowlist
• SSRF protection (block private IPs)
• Rate limiting

Git operations (code.repository.merge.v1):

• PR size limits
• Branch restrictions
• Review requirements

All policies at: https://aport.io/policy-packs

🔐 Security Model

What APort Protects

✅ Agent action security:

• Prompt injection (hook-based enforcement, not prompt-based)
• Malicious third-party skills
• Unauthorized commands
• Data exfiltration via files, messaging, web requests
• Resource exhaustion (rate/size limits)

Trust Model

APort operates at the application layer (between agent decision and tool execution).

You must trust:

• Your operating system (file permissions, process isolation)
• OpenClaw runtime (hooks execute correctly)
• APort code (open-source, verifiable)

Local mode additionally trusts:

• Filesystem integrity (passport not tampered)

API mode eliminates:

• Local passport tampering (fetched from API)
• Decision tampering (cryptographically signed)

Out of scope (OS/infrastructure security):

• File system compromise
• OpenClaw CVEs
• Network attacks (MITM, DNS poisoning)
• Supply chain attacks

This is standard for application-layer authorization (same model as OAuth, IAM, policy engines).

🎯 Use Cases

Protect against malicious skills:

• Install APort before adding community skills
• Every skill's tool calls are checked
• Malicious actions blocked before execution

Compliance and audit:

• Tamper-evident decision logs
• Court-admissible audit trail (API mode with Ed25519 signatures)
• SOC 2, HIPAA, SOX compliance support

Team deployments:

• Shared passport across systems (global suspend)
• Centralized policy updates
• Consistent enforcement

Air-gapped environments:

• Use local mode (zero network)
• All evaluation on-premise
• Self-hosted policy packs

📚 Documentation

APort Guardrails:

• QuickStart: OpenClaw Plugin
• Security Model & Trust Boundaries
• Hosted Passport Setup
• Tool/Policy Mapping
• Verification Methods (Local vs API)

OpenClaw:

• CLI: skills
• Skills Documentation
• Skills Config
• ClawHub

Security:

• SECURITY.md - Prompt injection, Cisco findings
• OAP Specification - Open Agent Passport standard

🤝 Support and Community

GitHub: aporthq/aport-agent-guardrails Website: aport.io Issues: GitHub Issues

Open-source: Apache 2.0 License Code review: All code publicly available for inspection

❓ FAQ

Q: Does this slow down my agent? A: Minimal overhead. API mode: ~60-100ms. Local mode: <300ms. Runs in parallel with agent thinking.

Q: Can I use this offline? A: Yes. Local mode works without network connectivity.

Q: What if I need custom policies? A: API mode: Pass custom policy JSON in request. Local mode: Edit bash script or use API mode.

Q: How do I suspend my agent? A: Local: Set passport status to "suspended". Hosted: Log in to aport.io and suspend (global effect).

Q: Is my data sent to APort? A: Local mode: No. API mode: Tool name + context only (no credentials, file contents, or env vars).

Q: Can the agent bypass this? A: No. Enforcement is in the platform hook (before_tool_call), not controllable by prompts.

Q: What happens if APort errors? A: Default: Tool blocked (fail-closed). Configurable via failClosed setting.

Made with 🛡️ by APort · Open-source on GitHub · Apache 2.0 License