rahul06x1/agentsense

This tool helps you understand what an AI agent actually did during a run. If your agent called the wrong tool, made an unexpected decision, or ran up a large bill and you are not sure why, this project records every step at the protocol level so you can look back at exactly what happened. It runs entirely on your own machine with no cloud account required. The core piece is an MCP proxy, a transparent go-between that sits in front of any MCP server your agent uses. You point your agent's configuration at the proxy instead of the real server, and from that moment on every tool call, its inputs and outputs, timing, and cost are captured without any change to your agent's code. The traces are stored in a local SQLite database as complete records with no fields left out. Two additional protections are built in. First, personal information is detected and replaced with consistent placeholder tokens before anything is ever written to the database, so sensitive data from users never lands in your logs. Second, the redaction uses a fixed process so that when you replay a trace later the placeholder values stay consistent and the replay makes sense. After capturing a run you can open a local web interface, built with FastAPI and plain HTML, to browse the trace as a tree of steps with a timeline view. The Compare tab lets you load two different traces side by side and see precisely where their decision sequences first split apart. There is also a live replay feature: you pick any captured trace, choose a different AI model, and the tool re-runs the session using the same recorded tool results rather than calling live services again. It then shows you a diff of what that model would have decided versus what your original agent did. This costs no extra API calls for the tool side. For agents you control directly, a Python SDK lets you record richer detail such as the reasoning steps and model calls that the proxy cannot see. The SDK writes to the same SQLite store and uses the same redaction path. The project is licensed under Apache 2.0.