Hoop

This platform, hoop.dev, provides runtime controls for AI agents that connect to databases, clusters and other infrastructure. A lightweight sidecar attaches to a resource and inspects every command an agent or engineer sends, masking sensitive data in responses, blocking dangerous operations and routing risky actions to a human for approval. Control happens in two layers. The first is deterministic: guardrail and masking rules the customer writes are evaluated in order and always return the same answer. The second is an agent that reads each statement, scores its intent and syntax risk, and judges what it would touch. Outcomes are pass, block, or route to a person in Slack or Teams with the exact command attached. Four controls are named: data masking, guardrails, a session analyzer and human in the loop approval. It works at the protocol layer across five protocols including PostgreSQL, MySQL, SSH, kubectl and gRPC, requires no code changes or new credentials, and the vendor cites under five milliseconds of added latency. It targets engineering and security teams that want to give coding agents such as Claude Code and Cursor access to production data without losing governance or an audit trail. An open source edition is free, MIT licensed and self hosted with no account, limited to one guardrail, one masking rule and one session analyzer rule per sidecar. Enterprise adds unlimited rules, a control plane and support, priced per identity, meaning each engineer, agent or service account, and quoted on request.