dbt open-sources Charts, a YAML dashboard language built for AI-generated analytics

What dbt announced

dbt Labs has released dbt Charts as an open-source project: a declarative language in which a full interactive dashboard is described in a single YAML file. According to a blog post by Dave Fowler on the dbt Charts site, the goal is to make dashboards produced through conversations with AI agents just as governable as dbt made SQL transformations.

Alongside the language, the company launched dbtCharts.com in public beta, a hosted platform that adds hosting, access control, a visual editor, version history and conversational analytics on top of the open format.

The problem it targets

Fowler frames the current choice in AI-assisted analytics as a fork. Left alone, an agent asked for one report tends to generate a sprawl of HTML, CSS, JavaScript, multiple chart libraries and eventually a React or Streamlit app. Tracing a number back to its source then means hopping across several languages and files — slow for a human reviewer, and costly in time and tokens for the agent on every revision. Traditional BI tools take the opposite route: copilots attached to UI-first products keep the agent inside sanctioned tooling, but only within what the UI happens to expose.

dbt Charts is pitched as a third path: the agent keeps working in code, but the output collapses into one structured file a person can actually read.

A continuation of BI's unbundling

The post situates the project in an essay by dbt Labs founder Tristan Handy titled "BI's Second Unbundling." In his account, early BI products were full-stack platforms covering ingestion, compute, transformation, semantics and visualization; from roughly 2015 to 2022 the modern data stack stripped out those infrastructure layers, leaving BI tools with visualization, sometimes semantics, access control and hosting. Charts stayed in the UI because pointing and clicking is faster than hand-writing config for most humans. Agents invert that preference: they are fluent in code, SQL and Git and clumsy in someone else's interface, so the chart layer is moving to where agents work.

How the language works

In a dbt Charts file, SQL expresses what data to fetch and YAML expresses how to present it. Markdown carries prose and Jinja supplies variables and macros, as in dbt itself. The post's minimal example defines one filter variable, one cumulative-count query and a single area chart, with a rows section laying out the board.

The CLI renders any board to static SVG, HTML, PNG, PDF or even the terminal, locally or in CI, and dct serve publishes a folder of boards as a site. Fowler says the language currently spans more than 1,100 configuration options across sixteen chart types plus charts composed from them. Styling cascades from theme to board to chart, boards can extend other boards, and a theme swaps in one line, so a house style or standard report is written once and inherited. The post also credits data graphic designer RJ Andrews with the visual system, positioning it as more crafted than the typical dashboard grid.

Tight coupling with dbt projects

The language stands alone, but the payoff comes with a dbt project. A charts directory sits beside models in the same Git repository, so a change to a model and the dashboards built on it ship together on one branch through one CI run. Queries reach models through ref(), resolved from the manifest, which means a renamed model or a missing column fails the offending pull request before anything reaches production. Validation runs with dbt parse followed by dct validate. Support for the dbt Semantic Layer, so boards can consume governed metrics instead of restating SQL, is planned and tracked in the project's first GitHub issue.

Feedback loops for agents

Because agents work best with tight feedback, the tooling strictly validates both the YAML and the SQL and adds visualization-level checks that flag problems before anyone sees the board. The post shows warnings such as a bar chart packing too many bands into the available pixels — with a suggested fix of rolling up to a coarser grain — and a table needing more width than its slot provides, suggesting fewer columns.

Why it matters

If analytics front ends are increasingly generated by chat rather than assembled by hand, the artifact that gets versioned, reviewed and trusted becomes the contract between humans and AI. dbt Charts bets that a compact declarative file is that artifact: diffable in Git, validated in CI and rendered deterministically to whatever format is needed. dbt pulled off the same move for the transformation layer and became its de facto standard; the company is openly aiming for a repeat at the visualization layer, the most visible piece BI tools have left. Whether that lands depends on adoption of the open language itself — and on whether the hosted dbtCharts.com platform, rather than the spec, becomes the real product.