BoltClaw

The security control panel for Claude Code skills and MCP servers

Why This Matters

Snyk's ToxicSkills study found prompt injection in 36% of scanned AI agent skills. Developers are installing skills and MCP servers from GitHub with no way to know whether they're safe - and every existing security tool assumes you already know what a YARA rule is. BoltClaw is the visual, plain-English alternative: scan before you install, see what your agent can access, fix misconfigurations before they become incidents.

How BoltClaw is Different

BoltClaw is the only tool that combines visual permission management, plain-English threat explanations, and guided setup - designed for the people who need security most: those who aren't security experts.

Architecture

┌──────────────────────────────────────────────────────────┐
│                      YOUR MACHINE                         │
│                                                           │
│  ┌─────────────────────────────────────────────────────┐ │
│  │              BoltClaw Dashboard                     │ │
│  │            (localhost:3000)                         │ │
│  │                                                     │ │
│  │  ┌──────────┐ ┌───────────┐ ┌────────────────────┐ │ │
│  │  │ Setup    │ │ Skill     │ │ Permission         │ │ │
│  │  │ Wizard   │ │ Scanner   │ │ Manager            │ │ │
│  │  └────┬─────┘ └─────┬─────┘ └──────────┬─────────┘ │ │
│  │       │              │                  │           │ │
│  │  ┌────▼──────────────▼──────────────────▼─────────┐ │ │
│  │  │              BoltClaw Config Engine             │ │ │
│  │  │         (reads/writes agent config)             │ │ │
│  │  └───────────────────┬─────────────────────────────┘ │ │
│  └──────────────────────│─────────────────────────────┘ │
│                         │                                │
│  ┌──────────────────────▼──────────────────────────────┐ │
│  │           Claude Code + MCP Servers                 │ │
│  │  Skills in .claude/skills/   MCP servers via JSON   │ │
│  │  Scanned before installation  Config hardened        │ │
│  └─────────────────────────────────────────────────────┘ │
└──────────────────────────────────────────────────────────┘

Prerequisites

• Node.js 22+ - nodejs.org
• Docker Desktop - docker.com/products/docker-desktop

Quick Start

# Install dependencies
npm install

# Start the agent + Ollama + Dashboard containers
npm run docker:up

# Optional: pull a local LLM model for agent testing
npm run docker:pull-model

# Run the Playwright test suite (requires containers running)
npm test

For local development without Docker:

npm run dev   # Starts dashboard at localhost:5173, proxies API to :3000

Authentication

BoltClaw's API requires a shared-secret token to prevent other processes on your machine from reading or modifying your agent config.

On first startup, BoltClaw generates a random token and saves it to ~/.openclaw/boltclaw-token. The token is printed to the terminal and also embedded in a clickable URL:

BoltClaw API token: abc123...
  Authenticated URL: http://localhost:3000?token=abc123...

Click the authenticated URL to open the dashboard. The token is stripped from the browser address bar automatically so it doesn't leak into history or referer headers.

To use a custom token, set the BOLTCLAW_API_TOKEN environment variable before starting BoltClaw.

Project Structure

boltclaw/
├── docker/               # Hardened agent container + docker-compose
│   ├── docker-compose.yml
│   ├── Dockerfile.dashboard
│   └── openclaw-defaults/ # Seed config
├── packages/
│   ├── config-engine/    # Parse, validate, backup, and write agent config + boltclaw.json
│   ├── skill-scanner/    # Static analysis - 15 patterns, 6 categories, risk scoring
│   ├── dashboard/        # React frontend + Express API (Setup Wizard, Scanner, Permissions, Audit Log)
│   └── mcp-server/       # MCP server exposing BoltClaw tools to Claude
├── tests/                # Playwright end-to-end tests (21 tests across 5 suites)
├── playwright.config.ts
├── package.json          # Monorepo root (npm workspaces)
└── CLAUDE.md             # AI assistant project context

Features

Setup Wizard

4-step guided flow with four security profiles: Lockdown, Balanced, Developer, and Migration Ready. Configure permissions with 3-way toggles (deny/ask/allow), and review your generated config with a security score before applying.

Skill Scanner

Scan Claude Code skills and MCP servers for threats before installation - paste a local directory path or a GitHub URL. Checks for 15 threat patterns across 6 categories: exfiltration (curl/wget, fetch, webhooks), prompt injection (instruction override, system prompt manipulation, tool output injection), obfuscation (base64, eval, hex encoding), permissions (sudo/privilege escalation, env variable access, undeclared capabilities), filesystem (sensitive file access, write operations), and execution (shell commands, browser automation). Each finding includes a plain-English "Why this matters" explanation. Risk scoring from 0-100 with levels: Safe, Caution, Warning, Danger.

Permission Dashboard

Visual grid showing current permission levels with colour-coded risk badges. Includes a circular security score gauge (A-F grading), Quick Fix buttons for risky settings, a findings list grouped by severity, and backup/restore UI.

Audit Log

Tracks every BoltClaw action: config reads, config writes, backup restores, and skill scans. Filterable by severity (Info/Warning/Danger), searchable by keyword, with expandable event details and a clear log button. Events auto-refresh every 10 seconds.

Config Engine

Reads and writes your agent config plus a boltclaw.json sidecar for BoltClaw-specific settings. Automatic backup before every change. Bidirectional mapping between BoltClaw toggles and config values. Auto-migration from old formats.

Testing

End-to-end tests (Playwright)

21 tests across 5 suites covering UI rendering, navigation flows, API validation, security input sanitisation, and config round-trips:

npm test          # Run all tests (headless)
npm run test:ui   # Run with Playwright UI for debugging

Unit tests

24 unit tests across two packages (dashboard API validation and skill-scanner analysis engine). No containers needed:

npm run test:unit

Docker Setup

3-container stack via docker-compose:

• Agent container - hardened with read-only filesystem, bridge network only, tmpfs for writable directories with correct uid/gid ownership
• Ollama - local LLM for testing without cloud dependencies
• BoltClaw Dashboard - serves the UI and API, includes git for GitHub URL scanning

Tech Stack

TypeScript, React, Express, Vite, Docker, Zod, Tailwind CSS, Playwright

Licence and Monetisation

BoltClaw is MIT licensed. Use it, fork it, ship it.

There is no paid tier, no Pro plan, and no cloud component. The codebase exists as a portfolio piece demonstrating threat modelling, static analysis design, and secure API design in the AI security space. Contributions welcome.