Most dashboards made by agents have only one user
Deepnote splits data apps into pages plus a notebook backend, letting Claude, Codex, and ChatGPT edit the same project. Beyond the demo, handoff and the second schema change are more worth watching.
Most dashboards written with agents have only one user: the person who built them. The charts are hard-coded snapshots embedded in an HTML file, and the database credentials sit locally. Send the file to a colleague, and they can't open it, or if they do, they can't connect to the warehouse. Jakub Jurovych put this awkwardness bluntly in Deepnote's launch.
Avi Chawla's tweet demonstrates a more practical state: Claude and ChatGPT editing the same live app, with data from the San Francisco dataset. In the video, the two models take turns editing the page, and the team sees an app connected to a live database, not an exported screenshot.
This time Deepnote has rebuilt data apps. Each app has a page that holds charts, filters, and interactions. Behind the page is a Deepnote notebook that executes SQL and Python. Database credentials stay behind the page. The team gets an inspectable, maintainable backend.
Specific capabilities include: queries can be rerun to refresh data and can also run on a schedule; supported OAuth integrations let logged-in viewers query with their own database permissions; version history can undo agent changes; colleagues access it through a shared link. Agents build and modify it through the Deepnote plugin or MCP. MCP can be understood as an interface that lets clients like Claude and ChatGPT call external tools.
Jakub's announcement video puts it more briefly: build it in Claude, a colleague picks it up in ChatGPT, open the link, and the data stays live.
The Deepnote blog fills in implementation details. You can start from Claude, Codex, Deepnote Cloud, or a local toolchain and end up in the same Deepnote project. The notebook backend runs SQL, Python, or R, and notebook inputs drive date pickers, filters, and dropdown menus. Apps are hosted in the workspace and connect to data sources such as Snowflake, Databricks, BigQuery, and ClickHouse.
Shared modules maintain metric definitions. For example, define revenue once, and both the revenue review and customer health apps can reuse it. Change the underlying calculation in one place, and the next time the imported notebook runs, it picks up the new value.
Scheduling lets notebooks refresh results before the weekly meeting or recompute account health every morning. Permissions are split into App user, View, and Edit: the first only uses the published interface, the second can open the project, and Edit can make changes. Per-user OAuth is supported, so queries run with the viewer's own warehouse permissions; private outputs keep each person's results separated. Public apps are controlled by the workspace, and admins can disable them.
What's really worth watching here is handoff. The cost of an agent producing a first version is already very low. The trouble starts when the second person opens the link. That second person needs the data to refresh, needs to know who maintains the SQL behind the charts, and needs to find someone to fix the join when the schema changes. In the demo, two models taking turns editing the same app looks smooth. In reality, a colleague first asks why the data didn't refresh, then asks who wrote this join.
Commenter neo put it more sharply: handoff is the real test. From Claude to ChatGPT, when the source table changes next month, who owns the SQL? Most vibe-coded dashboards die on the second schema change; the first build is actually the easy part.
Deepnote puts the notebook behind the page. The logic is reviewable, and data teammates can fix the bad join behind a chart. That solves the maintenance entry point. Ownership and concurrent editing still depend on team process. In the comments, someone asks who wins when two people edit the same app in different tools, and someone else asks whether the agent can catch bad SQL and bad cell output when checking for errors, or whether it mainly checks UI rendering. I didn't see a direct answer in the material. These questions are more worth watching than the feature list.
If your team is already using Claude Code or Codex to write internal tools, this direction is worth trying. Start with an app the team already needs, such as a weekly revenue dashboard, not with the demo. The trial app is here, the docs are here, and the examples are on GitHub.
Sources: Avi Chawla's tweet, Jakub Jurovych's tweet, Deepnote blog.
发布时间: 2026-11-19 23:27