Now available for OpenClaw

Control your AI agent.
Before it acts.

MantisBox is an autonomy governance layer that classifies, gates, and audits every tool call your AI agent makes. Stay in control without sacrificing capability.

MantisBox
$ mantisbox status
🦗 MantisBox Status
──────────────────────────────
Mode: execute-with-approval
Kill Switch: OFF

📊 Statistics
Total calls: 247
Allowed: 231
Blocked: 12
Pending: 4

Built for the AI agent ecosystem

⚠️

The Problem

AI agents can execute code, delete files, send emails, and access sensitive data. Once you give them tools, you lose visibility and control.

agent.execute("rm -rf /") Unchecked

The Solution

MantisBox intercepts every tool call, classifies the risk, and gates dangerous actions. You approve what matters, while safe actions flow through.

🦗 Blocked: destructive action Protected

Everything you need for safe AI autonomy

A complete governance layer that gives you control without limiting your agent's capabilities.

🎯

Smart Classification

Automatically categorizes every action as read-only, reversible, destructive, or privileged based on intelligent pattern matching.

🚦

Approval Gates

Risky actions wait for your approval. Safe actions flow through instantly. You set the thresholds.

📋

Complete Audit Trail

Every tool call is logged with full context. Know exactly what your agent did, when, and why.

🛑

Kill Switch

One command halts all agent execution instantly. Resume when you're ready. Peace of mind built in.

⚙️

Flexible Rules

Pattern-based rules let you always allow, always block, or require approval for specific tools and parameters.

🌐

Web Dashboard

Real-time web UI for approving actions, monitoring status, and reviewing audit logs from anywhere.

How MantisBox works

A simple but powerful pipeline that sits between your agent and its tools.

1

Intercept

MantisBox hooks into the tool execution pipeline. Every call passes through.

2

Classify

The action is analyzed and classified by risk level based on tool, parameters, and patterns.

3

Decide

Based on your autonomy mode, the action is allowed, blocked, or held for approval.

4

Execute

Approved actions proceed. Everything is logged to the audit trail.

Choose your autonomy level

Four modes to match your comfort level. Adjust anytime.

👁️

Observe

Log everything, execute nothing. Perfect for understanding what your agent wants to do.

Safest
📝

Propose

Every action requires approval. Full control over each decision.

Full Control
🚀

Fully Autonomous

Everything executes. Logging only. For when you trust your setup completely.

Maximum Speed

Get started in 60 seconds

MantisBox is an OpenClaw plugin. Install it, configure it, restart.

1

Install MantisBox

git clone https://github.com/JurassiCrafter/mantisbox
cd mantisbox
pnpm install && pnpm build
2

Add to OpenClaw config

// ~/.openclaw/openclaw.json
{
  "plugins": {
    "paths": ["/path/to/mantisbox"],
    "entries": {
      "mantisbox": {
        "enabled": true,
        "config": {
          "defaultMode": "execute-with-approval",
          "webPort": 7777
        }
      }
    }
  }
}
3

Restart & verify

openclaw gateway restart
openclaw mantisbox  # Check status

Ready to take control?

MantisBox is open source and free forever. Star us on GitHub to show support.