AI agents executing tool calls have no control point between model output and actual execution. Schema validation rejects malformed args — it can't decide whether DROP TABLE users should run right now.
Agentwall adds that layer. Classifies each call as safe, cautious, or destructive. Destructive calls require approval before running. Every attempt is logged as JSONL. Rollback hooks compensate if a session fails.
Python + TypeScript. Works with Anthropic, OpenAI, LangChain. Zero runtime deps.
