The AI-native BI tool built for agentic BI

An AI-native BI tool is one where an agent authors the same underlying object a human does, not a chat panel that clicks through someone else’s GUI on your behalf. dvt is built this way: every dashboard is a versioned JSON spec, and a human editing a panel in the browser, an agent posting a change over MCP, and a script calling the REST API all write the same document. That shared object model, not a bolted-on assistant, is what agentic BI actually requires.

Why most “AI BI” is a chat wrapper, not an AI-native BI tool

Most BI vendors added AI the same way in 2025 and 2026: a chat box next to the dashboard that translates a question into a query, or into a sequence of API calls that mimic clicking through the editor. The dashboard underneath is still an opaque blob in the vendor’s own database. An agent operating that surface can request a change, but it cannot read the dashboard as a document, diff it against the last version, or hand its edit to a second agent for review, because there is no document to hand.

dvt has no separate “AI surface” bolted onto the product. The spec (the JSON that describes a dashboard’s queries, layout, and visual encoding) is the object model for the editor, the REST API, and the MCP server alike. An agent that writes a spec through MCP produces the exact artifact a person produces by typing JSON or clicking through the live editor. One format, two kinds of author.

Agentic BI needs review, not just generation

Generating a dashboard once is the straightforward part. Agentic BI, the ongoing work of a team and its agents maintaining dashboards over months, needs a loop: a human has to see what an agent changed, revert it if it’s wrong, and leave feedback the agent can act on without a human re-typing the request from scratch.

dvt’s spec model makes that loop native. Every write is schema-validated before it saves, and any create or update can run in preview first, so an agent or a person sees the rendered result before it’s applied. Every version records whether a human, an agent, or the system made it, and which API key an agent acted through. A reviewer can leave a comment pinned to a specific panel, and an agent reading the open threads over MCP can make the requested edit, reply, and resolve the thread. None of that is possible when the object underneath is a GUI-state blob that nothing, human or model, can read directly.

Org knowledge agents read before they build

Being AI-native also means an agent isn’t starting from zero context on every build. dvt’s skills layer holds org-scoped, versioned documents: metric definitions, warehouse semantics, and the house rules for how your team builds charts. Agents read the relevant skills over MCP before authoring, so a dashboard an agent builds follows your conventions instead of generic defaults. Skills are written once, in-app or by an agent, and every agent with access reads the same version.

What an agent actually posts

A dvt spec is plain JSON. Here’s a minimal bar chart panel, the kind of document an agent constructs and the editor renders identically:

{ "id": "revenue-by-month", "title": "Revenue by month", "type": "chart:bar",
  "data": { "rows": [{ "month": "Jan", "revenue": 120 }, { "month": "Feb", "revenue": 150 },
                     { "month": "Mar", "revenue": 178 }, { "month": "Apr", "revenue": 168 }] },
  "spec": { "series": [{ "type": "bar", "barMaxWidth": 48 }] } }

Nothing here is a screenshot or a one-off render. It’s a validated document, diffable against the version before it, and reusable by the next agent or the next person who opens the dashboard.

Where this fits your stack

dvt connects to the sources analytics teams already query: Snowflake, Postgres, BigQuery, and Google Sheets. An agent queries your warehouse with credentials you already granted it, writes the spec, validates it, and applies it. dvt stores the spec, not a copy of your rows. Read the full spec reference or see the product’s version control and review tooling in depth.

Try the format yourself in the free Builder. No signup required. Ready to run it on your own warehouse? Get early access.