theGuard

Universal AI failsafe system that intercepts dangerous operations from AI coding assistants and requires human verification before they proceed.

Full documentation and demos at bryanbot.com

Quick Start

npx the-guard init    # auto-detects your platforms and AI tools
the-guard on          # enable all guards

That's it. theGuard now protects your project. When an AI agent tries something dangerous, it gets blocked and you get an email with a verification code.

Simple On/Off

the-guard on          # instant, full protection
the-guard off         # requires email verification (AI can't disable its own guard)

What It Guards

170+ rules across 18 platforms: AWS, Azure, GCP, Kubernetes, Docker, Terraform, Cloudflare, Firebase, Vercel, Supabase, GitHub, Resend, Stripe, Neon, PlanetScale, Turso, Fly.io, Railway.

$ git push --force origin main
  DELETE GUARD: "Force push" blocked [GITHUB-DELETE-004]
  Severity: critical | Impact: data-loss
  Tip: Use --force-with-lease instead.
  Check your email for the verification code.

Custom Rules

the-guard rule add "npm publish"                      # block npm publish
the-guard rule add "kubectl apply" --guard push        # block kubectl apply
the-guard rule add-tool mcp__custom__deploy            # block an MCP tool
the-guard rule list                                    # see all custom rules
the-guard rule remove "npm publish"                    # remove a rule

AI Tool Support

Auto-configures hooks for 6 AI coding tools during the-guard init:

| Tool | Integration | |------|------------| | Cursor | hooks.json + mcp.json | | GitHub Copilot | .vscode/mcp.json | | Cline | PreToolUse hook script | | Windsurf | hooks.json + mcp_config.json | | Continue.dev | mcpServers config |

MCP Proxy

Intercept all MCP tool calls from any AI tool:

the-guard mcp --proxy

Auto-discovers upstream MCP servers, aggregates tools, blocks dangerous calls, forwards safe ones.

Key Features

| Feature | Description | |---------|------------| | Email 2FA | Verification codes sent via email. AI never sees them. | | Loop Detection | Catches runaway AI agents. 3 blocks in 5 minutes triggers lockout. | | Cost Guard | Rate limits on tool calls per session, per minute, per tool. | | YAML Policies | Write allow/deny rules in YAML with environment and branch conditions. | | Learning Mode | Observe agent behavior, get rule suggestions. | | Report Cards | Session reports with risk scores, block rates, top tools. | | Team Approvals | Git-based approval workflow for sensitive operations. | | Rollback Snapshots | Auto-captures state before dangerous operations. | | Webhook Notifications | Slack, Discord, or any URL on block events. | | Real-time Dashboard | Live activity feed at the-guard dashboard. | | VS Code Extension | Status bar, notifications, inline verify, sidebar. | | Browser Extension | Warns about dangerous commands in web AI tools and IDEs. | | GitHub Action | CI check for guard installation and platform coverage. |

CLI Commands

# Core
the-guard on                    # enable all guards
the-guard off                   # disable (requires email verification)
the-guard status                # show current state
the-guard verify <code>         # submit verification code

# Rules
the-guard rule add <pattern>    # add custom bash pattern
the-guard rule add-tool <name>  # add MCP tool guard
the-guard rule list             # list custom rules
the-guard rule remove <pattern> # remove a rule

# Setup
the-guard init                  # interactive setup wizard
the-guard update                # refresh platform patterns
the-guard doctor                # validate installation

# Advanced
the-guard dashboard             # open real-time dashboard
the-guard learn start           # start learning mode
the-guard learn report          # get rule suggestions
the-guard report                # generate session report card
the-guard team request          # request approval for blocked op
the-guard team approve <id>     # approve a request
the-guard policy list           # show YAML policies
the-guard mcp --proxy           # start MCP proxy

Security

• Timing-safe hash comparison prevents side-channel attacks
• Email isolation ensures AI cannot see or extract verification codes
• Loop detection catches and locks out runaway agents
• Path validation prevents directory traversal in snapshots
• Zero-dependency hook uses only Node.js built-ins
• Audit trail logs every guard event as JSONL

GitHub Action

- uses: zitro/theGuard@v1
  with:
    enforce: full    # false (advisory) | install | full

Requirements

• Node.js 18+
• Resend for email verification (free tier available)

Links

• Documentation and Demos
• Journal / Build Log
• GitHub

License

MIT