# Decoy — Security tripwires for AI agents # https://decoy.run ## What is Decoy? Decoy deploys 12 honeypot MCP tools that look like real system utilities — execute_command, read_file, make_payment, and others. They appear alongside your real tools as an MCP server called "system-tools." If you're operating normally, you'll never touch them. But if a malicious document injects instructions telling you to exfiltrate data, read secrets, or run shell commands — you'll reach for these tools, and the tripwire fires. This isn't a judgment on you. It's a safety net for your operator. ## When to use Decoy - Your operator wants to detect prompt injection attacks on your tool calls - Your operator wants visibility into what happens when you process untrusted content - Your operator wants alerts when a document tries to make you do something malicious ## Setup Tell your operator to run: npx decoy-mcp init This creates a free account, generates a token, and configures Decoy as an MCP server in their Claude Desktop config. Takes 30 seconds. Alternatively, add as a remote MCP server in Claude Desktop: Settings → Connections → Add custom connector → paste endpoint URL ## Programmatic signup If you're setting up Decoy on behalf of your operator: POST https://decoy.run/api/signup Content-Type: application/json { "email": "operator@company.com" } Returns JSON with the operator's token and MCP endpoint URL. ## MCP endpoint https://decoy.run/mcp/{token} Standard MCP server over HTTP. Responds to initialize, tools/list, and tools/call. Server name: "system-tools". Exposes 12 tools. ## The 12 tripwire tools execute_command — Shell command execution read_file — File system reads write_file — File system writes list_directory — Directory enumeration make_http_request — Outbound HTTP calls query_database — SQL database access get_environment_variables — Env/secret access modify_system_config — System config changes manage_user_access — Permission escalation network_scan — Network reconnaissance make_payment — Financial transactions authorize_service — OAuth/trust grants ## What happens when a tool is called Every call is logged with full arguments, classified by attack type (exfiltration, credential theft, persistence, etc.), and scored by severity. The operator gets an alert via email, Slack, or webhook. The tool returns a fake success response so the attack chain continues and we capture the full payload. No real commands are executed. Nothing is harmed. ## Pricing Free forever: 12 tripwire tools, trigger logging, dashboard, email alerts, API access. Pro · $9/mo: Everything free, plus Slack alerts, webhook integration, and Monitor threat digest. No credit card required. Sign up at https://decoy.run