This guide takes you from one request to your AI agent to a colleague reading the answer in a data app: a small web app that queries your tenant as the person viewing it. You and your colleague open the same link, and each of you sees only the rows your own access allows. The steps are the same whether you type the commands or your agent runs them for you. Once Claude Code is signed in, the commands themselves take about two and a half minutes. Most of that is the first publish, which takes about two minutes. The rest of the time is your agent writing the app and you checking that it answers your question.

Before you start

Members can publish and install their own static apps, built with dip app new --runtime frontend, by default. You do not need an administrator for any step below, unless your administrator has turned own apps off for your role (see What if publishing or installing is refused?).

Build and share the app

1

Ask your agent for the dashboard

In Claude Code, describe what your team should see and say that they will open it. For example:
Build a dashboard of our open orders by region that my team can open.
The words “my team can open” matter. They tell your agent to build a Dippr data app, not a chat artifact, an HTML file or a spreadsheet of copied rows. Those would show everyone the rows you can see. Your agent says in one sentence that it is building a Dippr app, then looks up the tables it needs with the Dippr tools.
2

Your agent scaffolds the app

Your agent creates the app in a new folder, pointed at the catalog that holds your data:
It then writes the queries into the app. The app holds queries, never rows: every number on the page comes from a query run when someone opens it.The app is a static page. Dippr serves its built files itself and never runs your image on its servers. The page’s code runs in each visitor’s browser, and its queries go through your Dippr workspace, so each visitor sees only the rows their own access allows. See Static apps.
3

Look at the preview

Your agent runs dip app dev and gives you a local address to open. The preview queries your real engine as you, so it shows the rows your own access allows. Check that it answers the question you asked. Ask your agent for changes until it does.Nothing is shared yet. Only you can open the preview.
4

Publish the app

Your agent checks your computer and publishes:
dip doctor confirms Docker, buildx and that your login may publish. dip app push builds the app’s image on your computer and publishes it to your tenant. The first push of a new app takes the longest, because nothing is cached yet.
5

Confirm the install once

Your agent runs dip app install and shows you three things before anything runs:
  • the exact version it will install;
  • the data the app asks to read, with the reason the app gives, such as data:lakehouse:::read: “Reads the lakehouse catalog, subject to your grants, to render this app’s views.”;
  • how far to trust its code: “App code is trusted like a spreadsheet macro. It can read data allowed by each viewer’s access and this app’s declared access. Install only code you trust.”
Read them and answer yes or no. Your agent must ask you. Your sign-in, your existing access to the data and your original request do not count as approval.When your agent runs dip app install it cannot type yes for you, so the first run installs nothing. It stops and returns confirmation_required with a confirmation code. After you approve, your agent runs the install a second time with that code:
If the code is stale because the version or access changed, dip refuses and your agent shows you the new details to approve again. Once the second command runs, dip waits until the app is running and your agent shares the link only when the result is ready. The link comes with three lines like these:
These lines come from the access the installation actually holds. Your agent passes them on word for word.
6

Send your colleague the link

Send the link to your colleague. They open it, signed in as themselves, and the app queries the engine as them. If a row filter limits them to their region, they see their region and nothing else, from the same app and the same link. You do not filter anything inside the app; their own access does it.The top of the page reads Signed in as with their email (left out if their account has no email on record), and You see only the rows your access allows. Under each result, Data as of gives the time that query finished.The app is also on the Live Apps page in the console. Your colleague can open it there with Open.

Publish from Claude or ChatGPT on the web

In Claude or ChatGPT on the web there is no terminal for dip. Once the Dippr connector is added, ask your AI to build the page, then ask: “Publish this as a Dippr app.” The AI shows you the app’s name and what it will read, publishes it as you with the app_publish tool, and replies with the link. Publishing again under the same name updates the app. An app published this way is read-only: its page runs only its own files and reads only the catalogs it declares. Everyone who opens the link sees only the rows their own access allows.

Check that your colleague sees the right answer

Ask your colleague what they see, or compare it with what their access allows:
  • Their numbers should match a query they run themselves in Explore.
  • They should not see rows they cannot query themselves. If they do, the app is showing rows it read for someone else: check that its page runs its queries when it opens, rather than showing values copied into it.
  • If the app says “You don’t have access to this data — ask the app’s owner”, their access does not cover the query. Ask an administrator to grant it; the app cannot grant access.
Row filters belong to roles, and an administrator sets them in Settings → Roles & permissions. Administrators themselves are not limited by row filters, so check with a colleague who is not one. See Data grants, row filters and column masks.

How fresh is the data?

An app does not keep a copy of your data. It runs its queries each time someone opens it, so it shows what the table holds at that moment. The Data as of line under a result is the time that query finished.
  • Data from a connected source, such as Postgres or Snowflake, is read from that source each time.
  • Data you loaded from a file into a workspace database is a copy from the moment you loaded it. The app keeps showing that copy until you load the file again, for example with dip db load <database> --from orders.csv --table orders --mode replace, where <database> is the name of the database the app reads. Without a name, dip db load writes to your own personal database, which may not be the one the app reads. Nothing reloads it on a schedule.
An app that reads a workspace database needs that database’s catalog when you scaffold it (--catalog). Your colleague sees its rows only after you share the database with them.

What if publishing or installing is refused?

dip names the reason. Answer that reason, not a different one. You do not need the Data engineer role to publish or install your own static app. A Streamlit or Python app needs a Data engineer to publish it.

Build a data app

Scaffold, query, preview and publish an app in detail.

Install an app

What the install confirmation means, and how to change it later.

Use an installed app

What a colleague sees when they open the link, and what the errors mean.

Agent skills

How the skills steer your agent to a data app instead of copied rows.