Clawvisor

Clawvisor is an authorization and control plane for AI agents, delivered as an open source HTTP gateway that sits between an agent and the tools it touches. An agent declares a task (a purpose plus the tools that purpose needs), a human approves the scope once, and Clawvisor enforces it on every subsequent call. Real credentials stay in an encrypted server side vault (AES 256 GCM at rest) and are injected only at the moment of an approved call, so the agent holds only a scoped, short lived handle it can never reveal. A three layer access model combines hard restrictions, task scopes with silent auto execute, and a per request approval queue for out of scope actions. Every task is scored for blast radius by an LLM, and a full audit trail records each tool call, argument and decision, replayable and tied back to the authorising task, with token and tool spend attributed to the task and the person who approved it. It is model agnostic and, when self hosted, lets the operator choose the LLM provider (Anthropic, OpenAI, Ollama, Groq or Vertex AI).