Every change is reviewable, revertible, and attributed.
Because a dvt dashboard is a versioned spec, not a pile of clicks, the product ships the workflow that BI has always been missing: native version control with visual diffs, comments that a human and an AI can pass back and forth, and an org knowledge layer every AI build reads first.
The builder runs the spec format in your browser — no signup. Early access puts dvt on your own data; we reach out personally to set it up.
Watch dvt build a dashboard, then verify it.
One Claude Code session turns a live Postgres connection into a working dashboard, then traces a number on it back to the exact SQL, and through it to the table and column that produced it.
Every save is a version. Every version has a diff.
Every change, to a dashboard, a page, even a single element, gets an append-only revision. The History pane shows who changed what, with the change note and a before → after thumbnail for each version.
- Visual diffs. A pixel-level compare highlights exactly which regions moved: the AI's "did my edit do what I meant?" check, and yours.
- Identity-preserving restore. Roll back to any version; child elements keep their identity so later edits stay correct.
- Full attribution. Human, agent, or system, plus which key an agent used.
"Reworded the territory map legend endpoints for scannability (Top/Floor prefixes)."
Compare1d ago · 21:36 Jul 16REV 8* REV 11 "Softened the referral-network SMB nodes from red to emerald — red was reading as an alert state on the flows page."
Compare1d ago · 21:21 Jul 16REV 6* REV 10
Watch dvt turn a comment into a versioned diff.
A reviewer leaves a threaded comment pinned to a dashboard panel; an AI agent reads it over MCP, edits the spec, replies, and resolves the thread. The History pane logs the fix as an attributed revision, with the agent's change note citing the comment it resolved, and a spec-level diff that shows exactly what changed: the SQL, the series, the title.
Comment on a dashboard. Let Claude fix it.
Leave a threaded comment, pinned to a specific panel or the whole dashboard. An AI agent reads the open threads over MCP, makes the requested edits, replies, and resolves. The review loop humans expect from code, now for the dashboard itself.
- Anchored threads. A comment knows which panel it's about: by name, even after the panel is renamed.
- Two ways in. Ask Claude from the header's "Search or ask Claude…" field, or reach the review queue directly through dvt's MCP tools.
- Everyone can review. Comment and resolve sit at the viewer floor: reviewers don't need edit rights.
This should be MoM, not YoY — and can we sort segments by revenue descending?
via dvt-claude10:43Done — switched the delta to month-over-month and sorted segments by revenue. Preview looked right, applied as v14.
Teach dvt how your team builds, once.
Skills are org-scoped, versioned authoring docs: your metric definitions, warehouse semantics, and "how we build charts here." AI agents read the relevant ones before authoring, so every generated dashboard follows house rules, not generic defaults.
- Author once, apply everywhere. Humans write skills in-app or let an agent draft them: both write the same versioned record.
- Read at build time. Agents discover and load skills over MCP before they author.
- Different from the download. The public spec-authoring skill teaches any harness the format; org skills teach it your conventions.
Time-series charting conventions
House rules. Default to month buckets in the org timezone. Put the y-axis title above the axis, never rotated. Use the brand series palette; reserve red for churn and cost only.
Agents fetch this over MCP (dvt_skill_list) before authoring — so every generated dashboard follows the same conventions.
Watch an agent learn the house rules, then build by them.
A Claude Code session lists the org's skills over the dvt MCP, reads the ones that matter for a revenue build, profiles the warehouse, and builds a dashboard that follows them. On the human side, dvt shows the same skills as versioned, shared documents, down to a side-by-side diff of two versions.
More ways teams work in dvt
The workflow around the dashboard. Not just the picture.
Scheduled exports
Subscribe a dashboard, page, or single panel to recurring PDF/PNG exports: delivered on a schedule to email, Slack, Teams, or Google Chat.
White-label & dark mode
Set your product name, accent, and logo; pick a default appearance. Each viewer gets their own light / dark / system toggle over your org theme.
Folders & favorites
Organize dashboards into shareable folders with inherited access, and star the ones you live in so they surface first on your home screen.
Agent attribution
Every edit records whether a human, an agent, or the system made it, and which API key an agent acted through. Nothing an AI does is anonymous.
Preview before apply
Every write (human or AI) can be previewed first: validate the spec, compute the result, show the change, then persist. The AI shows its work before it makes it.
Built-in doc layer
Panels carry their own intent, assumptions, and the exact query that built them. The documentation travels inside the spec, not in a stale wiki.
Follow the build
dvt is in founding research. Leave your email and we'll reach out when there's something to see — no newsletters, no noise.