Noisegate - Differential-privacy gateway for untrusted AI agents

Show HN: Noisegate – a differential-privacy gateway for untrusted AI agents

Noisegate is a differential-privacy gateway enabling untrusted LLM agents to query sensitive data via the Model Context Protocol (MCP) with mathematical guarantees. It ensures no individual records leak, even if the agent is adversarial, by enforcing privacy in trusted code below the model. Validated by a runnable attack gallery and cross-checked against OpenDP, it offers three times the query utility through advanced accounting. Built with Python and DuckDB, it runs as an MCP server for Claude Desktop, providing a secure, deterministic trust boundary where the LLM proposes queries but enforces nothing.