We delivered a Chat for Looker application for a local print shop — one that lives inside the Looker instance they already run, answers plain-English questions from the dashboards and explores the team already trusts, and shows every query and chart it made to get there.
The shop already had the numbers. Orders, jobs, and customers were modeled in Looker, and the dashboards were good. The problem was the question that sits one step past a dashboard: why a tile moved, which customers were behind it, what it looks like broken out a different way. Answering those meant building a new explore, or waiting for the one person who knew how.
We delivered Chat, a Looker extension that puts a chat window inside their own instance. It installs from a few lines in a project manifest and shows up under Applications in the left sidebar, next to everything else they use. There is no new login and no new vendor. It is one React bundle running in Looker’s sandboxed frame, and every call it makes goes through Looker’s own extension SDK, so no credentials and no data leave the instance.
Most conversations start from a dashboard. Pick one from the sidebar, and its tiles become the context for the chat. Ask a question in plain English and it answers from the same models and explores the dashboard is built on, not a copy of the data somewhere else. Short updates show while it works. The answer arrives with its work folded above it: which fields it looked at, the SQL it ran, and the chart it built. Nothing is hidden. Every step opens up after the answer lands, so a number can be checked before anyone repeats it in a meeting.
When a table is not enough, it writes the chart. Using code generation, the agent builds interactive visualization tiles directly in the conversation, and a follow-up question refines them in place.
Access follows the rules the shop already set. Staff see only the dashboards, explores and agents their Looker permissions allow, with a second layer of control inside the app. Looker admins get a Settings tab where they tune the chat’s instructions, its suggested starter questions, and which data it is allowed to reach. When an answer comes back wrong, that is where it gets fixed, and it stays fixed for everyone.
It is also a starting point rather than a finished product. The same app can be themed to a brand, have its prompts shaped around a specific semantic model, and be shipped to production on any Looker instance. That work is scoped in one-hour blocks.