tophant-clawvault-mask-project

English | 中文

🎯 ClawVault is for you if

• ✅ You’re concerned about leaking personal private data when interacting with AI agents
• ✅ You want to prevent AI agents from accessing API keys, private files, and credentials
• ✅ You need to stop AI agents from mishandling sensitive or confidential files
• ✅ You want to keep logs of how AI interacts with your private data
• ✅ You need to detect AI injection attacks and dangerous commands

Activate your personal AI Vault:

• 1️⃣ Load private files
• 2️⃣ Set up and customize your secure storage
• 3️⃣ Create remote management skills

https://github.com/user-attachments/assets/2098b271-f844-4db5-b44a-f836e36d314c

Effect

https://github.com/user-attachments/assets/d580cfa1-8410-4095-90cb-3d693413a24e

Core Capabilities

1. Visual Monitoring

Users can configure their own "vault" and lock in Agents, Skills, credentials, and files they care about.
When someone touches these assets, the "Security Lobster" will notify you via IM: who touched what in your vault yesterday.

Technical Implementation:

• Event collection based on API gateway and file-side monitoring (invocation records, file access, change tracking)
• Supports periodic change notifications and real-time alerts

2. Atomic Control

Fine-grained control at the Agent level, using composable "atomic capabilities" as the smallest unit:

• Agent interaction and invocation policies
• Model routing, whitelists, and quota control
• Security detection (sensitive info recognition, credential detection, prompt injection protection, etc.)
• File access permission constraints

Users can combine these atomic capabilities like "building blocks" to create reusable policy configurations.

3. Generative Capabilities

Each "storage chamber" in the vault includes built-in basic security scenarios and allows users to add detection scenarios and Skills via natural language by mobilizing atomic capabilities.

Example:
Tell the system via chat interface:

For customer service Agent, if a user uploads a PDF containing 'contract',
it must first go through sensitive information desensitization,
and only GPT-4o-mini is allowed, with a single call limit of 2000 tokens.

The system will automatically generate and execute the corresponding policy rules.

---

✨ Features

• 🔍 Sensitive Data Detection — API keys, passwords, PII, credit cards, and 15+ pattern types
• 🛡️ Prompt Injection Defense — Block role hijacking, instruction override, data exfiltration
• ⚠️ Dangerous Command Guard — Intercept rm -rf, curl|bash, privilege escalation
• 🔄 Auto-Sanitization — Replace secrets with placeholders, restore on response
• 💰 Token Budget Control — Daily/monthly limits with cost alerts
• 📊 Real-time Dashboard — Web UI with Audit Log, Protection Monitor, per-agent config, and quick tests
• 🤖 Agent Integrations — OpenClaw session/runtime protection and Claude Code prompt/tool hooks

ClawVault combines a transparent proxy gateway, local agent hooks, and dashboard monitoring to protect AI provider traffic, OpenClaw sessions, and Claude Code prompt/tool activity.

🚀 Quick Start

https://github.com/user-attachments/assets/1f424172-8833-4ccc-b0d2-abf67c1758dd

Option 1: Install as OpenClaw Skill (Recommended)

# Install the dedicated installer from ClawHub
openclaw skills install tophant-clawvault-installer --version=0.2.9 --force

# Or install via clawhub CLI
clawhub install tophant-clawvault-installer --version 0.2.9

ClawHub: https://clawhub.ai/Martin2877/tophant-clawvault-installer

The installer skill provides AI-guided installation and management:

• /tophant-clawvault-installer install --mode quick - Quick setup
• /tophant-clawvault-installer health - Check status
• /tophant-clawvault-installer generate-rule "Block AWS credentials" - Create security rules
• /tophant-clawvault-installer test --category all - Run detection tests

See skills/tophant-clawvault-installer/ for skill documentation.

Operational commands after installation are provided by skills/tophant-clawvault-operator/.

Option 2: Install as Python Package

# Install
pip install -e .

# Start (proxy + dashboard)
clawvault start

# Scan text
clawvault scan "password=[PASSWORD_EXAMPLE] key=[SYNTHETIC_API_KEY]"

# Interactive demo
clawvault demo

🚀 Deploy to Server

# One command: pack, upload, install
./scripts/deploy.sh <server-ip> root

# On server: setup integration + start
./scripts/setup.sh
./scripts/start.sh

📜 Scripts

🏗️ Architecture

    OpenClaw
       │
       ▼
┌─────────────────────────────────┐
│    ClawVault (Security Vault)   │
├─────────────────────────────────┤
│ Gateway Module                  │
│  • Transparent Proxy  :8765     │
│  • Traffic Interception         │
├─────────────────────────────────┤
│ Detection Engine                │
│  • Sensitive data               │
│  • Injection patterns           │
│  • Dangerous commands           │
├─────────────────────────────────┤
│ Guard / Sanitizer               │
│  • Allow / Block / Sanitize     │
├─────────────────────────────────┤
│ Audit + Monitor                 │
│  • SQLite storage               │
│  • Token budget tracking        │
├─────────────────────────────────┤
│ Dashboard                       │
│  • Web UI :8766                 │
│  • Agent config & tests         │
└─────────────────────────────────┘

⚙️ Configuration

# ~/.ClawVault/config.yaml
proxy:
  port: 8765
  intercept_hosts: ["api.openai.com", "api.anthropic.com"]

guard:
  mode: "interactive"  # interactive | strict | permissive

monitor:
  daily_token_budget: 50000

OpenClaw OpenAI-Compatible Adapter

For OpenClaw deployments, prefer the direct OpenAI-compatible adapter over a transparent proxy when you want @clawvault sanitize prompts to continue to the model after local sanitization.

OpenClaw provider configuration:

base_url = http://127.0.0.1:8766/v1

Keep the existing OpenClaw model and provider API key unchanged. Configure ClawVault to forward adapter traffic to the real upstream provider:

# ~/.ClawVault/config.yaml
provider_adapter:
  enabled: true
  upstream_base_url: https://api.example.com/v1
  request_timeout_seconds: 60.0

Do not set `provider_adapter.up