This is an exactly-once execution guard that prevents AI agents from firing duplicate tool calls when Claude retries after errors or crashes. It exposes a /claim endpoint that returns PROCEED on the first call for a given request_id and SKIP on any repeat, blocking double charges, duplicate emails, or repeated orders. The server runs on Railway with pay-per-call pricing via x402 ($0.001 per claim), plus a free /claim/test endpoint for integration testing. Also includes a local SQLite guard you can embed directly in Python agents. The repo shows a real trading bot that prevented duplicate buy orders across 23 executions using stable keys derived from symbol, side, quantity, and bar timestamp. Useful when you're building agents that interact with payment APIs, email services, webhooks, or any broker where firing the same action twice causes real problems.
claude mcp add --transport stdio azender1-safeagent uvx safeagent