Prototype (AG-336). The AI security server endpoint for Anthropic Inference hooks. Anthropic POSTs each governed frame here and waits for an allow or deny verdict. Configure it in claude.ai with this URL and a custom Authorization: Bearer <Opal API key> request header.

tool_call frames are evaluated against the organization's runtime policy rules. One verdict covers every tool call in the frame. Every other frame type, including prompt, is allowed without inspection. Any non-200 response is a webhook failure, so the organization's Anthropic failure-handling setting decides the outcome.