Blog · · Dave Fowler

Claude Dashboards you can Govern

Add the dbt Charts connector and your agent gets more answers right and saves dashboards to Git, where your data team governs them.

Claude can build dashboards now. Ask for one and it appears in the conversation a minute later, and anyone at your company can do it.

The question is what happens next. The dashboard lives in one person’s chat. Nobody reviewed the query, nobody checked whether that person should see the numbers, and the data team hears about it when a figure from it is quoted in a meeting. The usual answer is to treat the chat as a sketchpad: prototype the dashboard there, then rebuild it in a BI tool before anyone relies on it.

Self-serve analytics finally arrived. It shouldn’t need a rebuild to be trusted. The dbt Charts connector gives you both: anyone can ask for a dashboard, and the data team still owns every number in it.

Dashboards as (governable) code

Today we’re launching the dbt Charts connector for Claude, ChatGPT, Codex, and any other agent tool that speaks MCP. Add it to the assistant your company already uses, and the same request returns a dashboard written in dbt Charts: drawn in the conversation, running on your warehouse, and stored as one YAML file a person can read.

For the person asking

  • Ask in the assistant you already use: Claude, ChatGPT, Codex, and anything else that speaks MCP.
  • Get a live dashboard in a minute, with no SQL, no install, and no ticket.

For the data team

  • Every call is checked against that person’s own permissions.
  • Every saved dashboard is a file in your repo. Turn on branch mode and each one is a pull request, reviewed before it goes live.
  • Nothing to rebuild, and nothing ungoverned to chase down.

A ChatGPT conversation using the dbt Charts connector: asked for Dundersign’s top five salespeople, it renders a bar chart led by Dwight Schrute at $148,608; asked for the main sales dashboard, it opens the Bookings & Recurring Revenue board with cumulative won revenue, win rate, ARR by plan, net new subscriptions, key figures, and a monthly table, and a tooltip on the ARR pie

The person asking sees a chart, and never YAML or SQL. They can ask for changes, click into the editor to adjust a label, or ask the assistant to save it to a folder. The data team sees a file in the repo.

Less for the data team to review

Governance only works if someone can read what the assistant built.

BoardBench is our benchmark for agent-built dashboards. An agent gets a written brief and a dataset, builds a dashboard, then handles a styling request and a set of review notes. Left alone, it writes HTML, CSS, JavaScript, and a script to prepare the data. In dbt Charts it is one YAML file. Across our last six rounds, about 160 runs a side on 14 held-out datasets, Codex wrote 7,800 output tokens with dbt Charts and 16,600 building a dashboard its own way: less than half as much for the data team to read.

And it gets more answers right

Self-serve only works if the answers are right. Give the assistant the data team’s project and it stops guessing.

AskBench is our benchmark for data questions. An agent gets a question a real team asked of its warehouse and has to come back with the right number. We run each model twice: once with only the warehouse, and once with a dbt Charts project in reach.

Percent of real data questions answered right. Gray is the agent on the bare warehouse; orange is the same model with a dbt Charts project in reach.AskBench: right answers with and without dbt ChartsPercent of real data questions answered right. Gray is the agent on the bare warehouse; orange is the samemodel with a dbt Charts project in reach.+42 pts+24 pts+27 pts+38 pts+18 pts+13 pts30406080100Claude HaikuClaude SonnetClaude OpusGPT-6 LunaGPT-6.1 TerraGPT-6.1 Sol
See the YAMLa example
charts:
  right_answers:
    query: askbench
    type: bar
    title: "AskBench: right answers with and without dbt Charts"
    subtitle: Percent of real data questions answered right. Gray is the agent on the bare warehouse; orange is the same model with a dbt Charts project in reach.
    x: model
    y: with_charts
    y_start: bare
    style:
      orientation: horizontal
      legend:
        visible: false
      marks:
        bar:
          band_width: 0.03
          color: "#b5b5b5"
      axis_y:
        grid:
          torn_baseline:
            enabled: true
        scale:
          continuous:
            domain: [30, 100]
    layers:
      - type: scatter
        y: bare
        label: Bare warehouse
        style:
          marks:
            point:
              size: 140
              color: "#9a9a9a"
      - type: scatter
        y: with_charts
        label: With dbt Charts
        style:
          marks:
            point:
              size: 140
              color: "#FE6703"
              labels:
                visible: true
                field: gain
                position: right
                dx: 10

We ran six Claude and OpenAI models on the same 197 questions, and every one got more answers right with the project. Claude Opus went from 61% to 89%, Claude Haiku from 37% to 78%, and GPT-6.1 Sol from 81% to 94%. The agent stops guessing at which table holds revenue and reads the query a dashboard already runs.

The warehouse does less work too: on GPT-6 Sol the project scanned 37% fewer bytes.

Saved next to your dbt models

A dashboard sketched in a chat and rebuilt in a BI tool is two dashboards: the one someone asked for and the one the data team maintains. They drift apart, and the second lives in a vendor’s format that only that vendor’s product can edit.

With dbt Charts there is one. The dashboard the assistant drew is a YAML file in your dbt project, beside the models it reads:

my_dbt_project/
  dbt_project.yml
  profiles.yml
  models/
  charts/
    sales/
      bookings_and_recurring_revenue.yml
      top_salespeople.yml
    themes/
      newsroom.yml

A person can read it, a reviewer can see what changed, and a renamed column fails the pull request that renames it. dbt Charts is open source, so if you ever stop using it, you keep a folder of readable files and the SQL inside them.

What the data team keeps

  • It signs in as you. There is no service account. Every call is checked against that person’s own access to the dashboard, the project, and the data source, so the assistant can reach at most what the person already can.
  • It carries queries, not data. The file holds the SQL, so it runs against the warehouse each time it opens, and each query is labeled in your warehouse’s job history.
  • It saves only when asked, and it saves to Git. What the assistant builds is a private draft. Saving it is a commit in your repo, and with branch mode on, a pull request reviewed as a picture before it goes live.
  • It can be switched off. Account Settings lists every assistant with access. Disconnect one and its next call fails.

Add it to your tools

The connector is a remote server at one URL, so there is nothing to install. On Claude’s Team and Enterprise plans an owner adds it once and everyone in the organization gets a Connect button.

  • Claude, on the web, desktop, and mobile: add the connector.
  • ChatGPT: add the connector by URL and sign in.
  • Codex: codex mcp add dbt-charts --url https://dbtcharts.com/mcp.
  • Claude Code and Cowork: the dbt Charts plugin bundles the connector with the skills for building, styling, and reviewing dashboards.
  • Anything else that speaks MCP: point it at https://dbtcharts.com/mcp and sign in.

Every tool and permission is listed in the connector docs.

Try it

Your business users get answers in the tools they already use. Your data team keeps the models, the reviews, and the final say.

The connector is free to add, and it works with the free workspace. There’s no license to buy and no demo to book.

Add it, ask for a dashboard, and tell us what it got wrong. 

Newsletter

Get the next post by email

Notes from building dbt Charts: the language, the renderer, and what we learn shipping BI as code.

Prefer a feed? Subscribe by RSS.