TradeDog - The Datadog for AI Trading Agents

Everyone is building trading bots that attack. Nobody is building the immune system around them. Thousands of bots run on brokerage APIs today with no insurance, no supervision, and no one watching at 3am — the October 2025 flash crash liquidated $19.3B precisely because herding algorithms had no protection. TradeDog is the guardian layer: Datadog watches your servers, TradeDog watches your trading agents. WATCH — A deterministic watcher polls the account every 60 seconds via EventBridge: stuck orders, order bursts (rogue bots), daily-loss breaches, silent bots, unprotected exposure. FIX — Claude (Anthropic SDK) diagnoses each anomaly and proposes exactly one action from a closed menu; deterministic Hands validate it against hard risk gates and execute through the Trading API. Emergencies skip the LLM entirely: kill switch first, explanation after. Fail-closed by design. INSURE — The dog builds near-zero-cost collars (protective put + covered call) as single multi-leg (Level 3) orders, pricing strikes from live option chain quotes so the short call finances the long put. Both of our positions got collared autonomously on day one — real fills, no human involved. HERD RADAR — TradeDog consumes our own Agent Sentiment Index (RSoft Sentiment, computed from ERC-8004 verified on-chain agents, live on mainnet). When the AI herd crowds one side, the dog warns you before the stampede. TALK — Every alert reaches the owner as a plain-language WhatsApp message, and the owner can chat back or use slash commands: /status, /protect, /report, /kill. Every event, LLM proposal, and execution lands in an append-only audited ledger (Supabase). Built solo in 3 days on Alpaca's Trading API, MCP server, and CLI, with a Datadog-style Next.js dashboard, AWS Lambda + EventBridge, and Claude as the brain that proposes but never pulls the trigger.