AI Agent Security Engine — Block dangerous operations, audit all behaviors, protect sensitive data.
Built for OpenClaw · Also supports Cursor / Windsurf / Claude
AgentGuard is a security layer for AI agents. It intercepts dangerous commands, prevents data leaks, and maintains a tamper-proof audit log of everything your agent does.
| Layer | Name | Description |
|---|---|---|
| L1 | Gateway Lockdown | Blocks native exec/write/edit tools at the gateway level — agents physically cannot bypass |
| L2 | Rules Engine | Command blacklists, file path controls, domain whitelists |
| L3 | Redaction Engine | Auto-filters API keys, tokens, SSH keys and 15+ other sensitive data types (input + output) |
| L4 | Audit Log | All operations logged to local SQLite — viewable in Dashboard, immutable |
| Platform | Status |
|---|---|
| OpenClaw | ✅ Full support |
| Cursor | ✅ Supported |
| Windsurf | ✅ Supported |
| Claude | ✅ Supported |
# Install AgentGuard daemon
curl -fsSL https://www.agentguard.site/download/install.sh | sh
# Start daemon
agentguard daemon start| Tool | Replaces | Security |
|---|---|---|
ag_execute_command |
exec / process |
Command blacklist + dangerous command interception + output redaction |
ag_read_file |
read |
Sensitive path blocking (.ssh/, keychain, browser data) + content redaction |
ag_write_file |
write / edit |
Path access control + write content inspection |
ag_list_directory |
read (directory) |
Directory access control |
ag_http_request |
network tools | Domain whitelist + data exfiltration prevention |
| Tool | Purpose |
|---|---|
ag_skill_check |
Check if a skill/plugin is in the security whitelist |
ag_status |
View engine status: mode / audit stats / panic state |
| Tool | Purpose |
|---|---|
ag_panic |
🚨 Emergency stop — immediately reject all subsequent operations |
ag_resume |
Resume normal operation |
- enforce — Violations are blocked immediately
- supervised — Suspicious operations pause and await user approval
- permissive — Audit log only, no blocking
View the full audit log at: http://127.0.0.1:19821
Features: real-time operation timeline · audit charts · rule configuration · one-click panic
- 🌐 Website: agentguard.site
- 📦 Skills: skills/agentguard/SKILL.md
AgentGuard 是一个 AI Agent 安全引擎，通过四层防护机制确保 Agent 操作安全可控：
- L1 网关封锁 — 在 Gateway 层物理禁用危险原生工具
- L2 规则引擎 — 命令黑白名单、路径控制、域名白名单
- L3 脱敏引擎 — 自动过滤 API Key、Token 等 15 类敏感信息
- L4 审计日志 — 所有操作记录本地 SQLite，不可篡改
支持平台：OpenClaw · Cursor · Windsurf · Claude