The first time I ran an AI agent overnight I woke up to a $47 bill and no clear idea what it had actually done. The agent had finished its task — sort of — but somewhere in the middle it had looped on a subtask, made the same API call sixteen times, and burned through tokens I hadn't budgeted for. I had no log, no trace, no replay. Just a number on a dashboard and a vague sense of unease.
I looked around for something that would show me, in real time, what my agents were doing. The tools I found were all cloud-based: pipe your agent's traffic through our servers, sign up for an account, and we'll show you a pretty dashboard. That meant handing off my prompts, my responses, and my architecture to a third party whose business model I didn't fully understand. For production agents working with sensitive data, that felt like exactly the wrong trade.
AgentWatch is what I actually wanted: a local CLI process that intercepts nothing, sits beside your agent as it runs, and gives you full cost and behavior visibility without a single byte leaving your machine. It's not a SaaS. It's a tool — yours, running locally, answering only to you.