Claude Dashboards: how to build one, and check every number on it

Type "a dashboard of open pipeline by stage and owner, from Salesforce" and a minute later there are charts. Click any one and you see the SQL behind it, plus the time its data was last refreshed. That's Claude Dashboards, launched in beta on 8 October 2026, and the visible query is the feature that matters most.

What Anthropic documents, as of 22 October 2026:

Anthropic is clear about the job it's for. Its own FAQ answers "Does Claude Dashboards replace my BI and analytics tools?" with: "Claude Dashboards is made for quick questions about your data." We agree, and that's the frame for the rest of this page: a fast first draft whose every number you can trace, not a reporting system.

Fifteen minutes, most of it spent on step 2.

1. Check the plan. Pro or above. On Enterprise, ask an owner to enable Dashboards under Organization settings → Artifacts (Anthropic Help Centre).
2. Connect the warehouse read-only. Add its official MCP server as a custom connector: Customize → Connectors → Add custom connector on Pro and Max, or an Owner adds it under Organization settings → Connectors on Team and Enterprise (custom connectors). Snowflake's managed server runs SQL as SELECT only by default (Snowflake docs); BigQuery's tools are read-only apart from execute_sql, which an IAM deny policy blocks (Google Cloud docs). Our Snowflake MCP and BigQuery MCP guides walk through both.
3. Start the chat where your definitions are. Anthropic says you can ask in any chat. We ask in a Claude project whose instructions define the metrics, because a dashboard inherits whatever Claude decides "active" means.
4. Ask for it. Either select "Output" in the message box and choose "Dashboards", or open the Artifacts tab and pick a Dashboards template. Name the table or app; the help article says that helps. An illustrative request:

Build a dashboard from analytics.fct_subscriptions:
1. Active customers by plan, weekly, last 12 fiscal weeks.
2. MRR by month, last 12 months, in GBP.
3. New vs churned customers by month.
Use the definitions in this project's instructions and put the
definition used in each chart's subtitle. Exclude internal
accounts (is_internal = true).

1. Open every chart's query before you look at the shapes. Anthropic's own advice: open the query, and if it isn't counting what you meant, tell Claude what to change.
2. Refine in chat, one change at a time ("add a plan filter", "use fiscal months").
3. Share from the dashboard's Share button once the queries are right.

From a spreadsheet instead of a warehouse

Anthropic's Artifacts page says a dashboard can start from "a spreadsheet, a report, or a live connection". The help article covers only warehouses and connected apps, and neither page says how a dashboard built from an uploaded file refreshes. Treat one as a snapshot of the file you uploaded, and if the file is the real source of truth, Claude data analysis covers how Claude handles uploads.

Snapshot or live, a dashboard is only as right as the queries behind it. With Count connected to Claude, each number comes with its reasoning chain: the source, the calculation, or the assumption nobody tested. Check the tiles before you share the link.

Anthropic's launch article says Claude builds the dashboard "and keeps it current as the data changes". Its Artifacts page gives a Salesforce example that "updates as deals move". So by Anthropic's account, a dashboard on a connector is live rather than a picture.

What isn't documented: how often it refreshes, whether you can schedule a refresh, or what happens when a connector's sign-in has expired. The help article and the Artifacts help page both say only that each chart shows "when it was last refreshed". That timestamp is the thing to read before you quote a number in a meeting.

Two documented behaviours change what a shared dashboard shows. "Everyone connects their own apps, even when using a shared or published artifact," says Anthropic's Artifacts help page, and viewers' usage "counts against each person's own plan limits". The upside is that a dashboard can't leak data its viewer couldn't already query. The catch: two colleagues can open the same dashboard and see different numbers if their warehouse permissions differ. Sensible for privacy. Confusing in a board meeting.

A Claude dashboard is a kind of artifact, so it helps to know what's different.

Sources: Anthropic Help Centre and Artifacts page, October 2026.

The column that matters is the last row but one. In a BI tool, change the definition of churn once and every chart picks it up. In Claude, a definition lives in one project's instructions, and a dashboard made in another chat may use a different one. That's why we'd build ad hoc and exploratory dashboards in Claude and promote anything people check every week into the BI tool, which is where Anthropic points you too: it lets you send a dashboard on to Hex, Omni, Sigma and others. There's a fuller comparison of AI dashboard tools in AI dashboards and of artifacts in general in Claude artifacts.

Limits worth knowing

• It's a beta. Behaviour and plan availability may change; we'll update this page when Anthropic does.
• Usage limits apply, to you and to each viewer.
• Connector limits flow through. BigQuery's MCP server, for one, times out after three minutes and caps results at 3,000 rows (Google Cloud docs), so a chart over a large table may be summarising less than you think.
• Approval-gated tools are out. "Connector tools that need approval for each action aren't available to artifacts."
• No documented size or row limits for Dashboards itself, and no documented refresh schedule.

Twelve tiles look like one thing: a dashboard. They're twelve claims, each resting on a table choice, a join, a filter and a definition. The chart renders whether or not those were right, and nothing on screen tells you which tile's WHERE clause quietly dropped a quarter of your customers.

The evidence is why we'd insist on checking. On Spider 2.0, built from real enterprise databases, an o1-preview agent solved 21.3% of tasks, against 91.2% on the academic Spider 1.0 (Lei et al., ICLR 2025). Agents have improved since, but the failure mode hasn't changed: SQL that runs and answers a slightly different question. More in text to SQL.

Claude Dashboards makes the check unusually cheap, because every chart already carries its query. Use it.

A two-minute check before you share

1. Pick the two tiles the meeting will argue about. Usually the headline number and the trend.
2. Open each query and read three things: the table, the date filter, and any WHERE clause that excludes rows (test accounts, trials, refunds).
3. Compare the definition in the query with the one in your project instructions or glossary.
4. Tie one number to a source you trust: last month's board pack, the billing system, a query you wrote yourself.
5. Read the refresh time on every tile you'll quote.

If you'd rather have this done claim by claim, Count connects to Claude over MCP: ask it to audit the dashboard's answer and it returns a canvas where each claim is labelled as a citation, an assumption, model knowledge or a hallucination, with the reasoning chain behind it. The general method is in how to check AI answers.

Is Claude Dashboards free?

Does a Claude dashboard update automatically?

Can I make a Claude dashboard from an Excel or CSV file?

Where is the Claude API usage dashboard?

Does Claude Dashboards replace a BI tool?