A browser co-pilot with an approval gate, and the training to run it yourself
Built for our own operation and shown here as a capability example. The 4-hour training was delivered to Alex, an auction-platform operator. First name only, used with permission; no NDA, low confidentiality risk.
Our own team’s internal product. Training client: Alex, an individual auctioneer and online auction-platform operator.
Problem
Our own operation ran the same repetitive web tasks by hand, one at a time, across the platforms we work in. The obvious fix, a script that logs in and clicks, was the thing we could not safely use: hand a script your live accounts and it keeps clicking even when a page changes or something looks wrong, with no judgment about when to stop and ask.
Approach
We built a self-contained Chrome Co-Pilot Kit that lets an AI assistant work inside a private copy of your own signed-in browser, learn a task by watching you do it once, then repeat it the cheapest reliable way it can, from a direct behind-the-scenes connection down to clicking through the pages step by step the way a person would.
Nothing that changes data or account settings runs without a person approving it first. The delivery was documented honestly, as a plain map of what works today versus what is still a rough framework, so the operator knows exactly what to rely on.
Alongside the build, Alex, an auction-platform operator, took the 4-hour hands-on training kickoff: how to direct AI to do the work (agentic development), how to keep it in bounds (governance, meaning spending limits, budgets, keeping a person in the loop, and holding approvals for later), and the 11-step process we use to take work from idea to shipped, with a take-home assignment to build his own interface for the tool.
What shipped
- A working core: driving the browser, learning a task by watching what it does once, upgrading that task to the fastest method that still works reliably, and the human-approval gate, all validated end to end against a live site with 51 automated tests.
- A documented map of what is proven versus what is still a rough framework (multi-panel scheduling and a link between the engine and the browser extension are flagged as not yet built).
- A 4-hour training course delivered, covering how one lead AI coordinates several working ones, how to keep them governed, and the 11-step path from idea to shipped, plus a take-home build assignment.
Verified numbers
- 51automated testsvalidated end to end on a live site
- 4 hrshands-on training delivered
- 11steps in the build process taught
- 1human approval required before anything changes data or settings
What this means for you
If you need a tool built and the ability to run it yourself, both are delivered together: a working system with an honest account of what still needs finishing, and training built so your team can run and extend it without us.