Show HN: AgentWatch – Local MCP proxy that catches agents leaking your data (github.com)

AgentWatch has launched a local tool designed to monitor AI agents (such as Claude Code and Cursor) and prevent potential data leaks when these agents interact with untrusted sources. By placing itself as a proxy between the agent and its Model Context Protocol (MCP) server, AgentWatch tracks all tool calls in real-time, flagging any instances where data fetched from an untrusted source is sent out externally. This transparency addresses the "black box" nature of AI agents, enabling users to understand and mitigate risks associated with accidental data leaks. The significance of AgentWatch lies in its ability to enhance data security within AI applications, particularly as reliance on such agents grows. Utilizing a taint tracking technique, the tool identifies when data from untrusted origins is passed through to outbound calls, offering user-friendly explanations for each risk encountered. The dashboard provides live updates, session stats, and risk indicators, all while keeping the system fully local—ensuring no data is transmitted externally. While version 0.1 is an observatory tool with known limitations, including a lack of granularity in identifying transformed data leaks, future updates aim to introduce enforcement capabilities and broader support for various MCP transports.