hoophq/hoop

What it solves

Hoop provides runtime safety controls for AI agents that need access to sensitive data and infrastructure. It prevents agents from executing dangerous operations (like dropping database tables) and ensures sensitive information (like emails) is redacted before it reaches the agent, without requiring changes to the agent's code or prompts.

How it works

Hoop operates as an open-source sidecar binary that sits between the AI agent and the resource (such as a database). The agent connects to the sidecar instead of the resource directly. The sidecar inspects the wire protocol traffic in real-time to apply four types of controls:

• Data Masking: Rewrites sensitive values in the response in-memory before they reach the client.
• Guardrails: Uses a deny list to block destructive statements from reaching the database.
• Session Analyzer: Uses an agent to score the intent and syntax of actions for risk.
• Reviews: Escalates high-risk operations to a human for approval.

It supports protocols including PostgreSQL, Microsoft SQL Server, and HTTP.

Who it’s for

Developers and organizations deploying AI agents into production who need to govern the connection between the agent and their data resources to prevent accidental data loss or privacy leaks.

Highlights

• Zero-config for agents: No SDKs, prompt changes, or agent-side configuration required; only the connection port changes.
• Protocol-level control: Works at the wire level, making it independent of the specific LLM framework or MCP interface used.
• Multi-protocol support: Compatible with PostgreSQL, MSSQL, and HTTP.
• Sidecar architecture: Runs alongside the resource rather than replacing it, preserving existing identity and sandbox stacks.