datatomcp

Totals, counts, averages, min and max, optionally grouped. Metrics: count, count_distinct, sum, avg, min, max, or { metric: name } for one of the team's approved definitions (describe_dataset lists them). order_by may use group columns or metric names. where, group_by and metrics may use columns of linked datasets as alias.column (see describe_dataset links). sum, avg and count of a linked column count each linked row once per group (a customer's credit limit once, however many orders), so revenue and target by country can come from one call.

Draw a chart from a query on a dataset: the query runs like aggregate (or query_rows when it has select), then result columns map onto the chart. x is the group_by column for the axis or categories; y lists metric names (or columns); series splits line, area, columns and bar into one series per value. grain (day, week, month, quarter, year) groups a date column in group_by by that period. mode draws columns and bars stacked, grouped or as 100%. variance, bullet and scorecard take two y columns (actual, then plan or target); candlestick takes four (open, high, low, close). Measures defined as lower is better (costs) are coloured the right way round. Write the title as the takeaway ("Asia Pacific grew fastest in 2026"), not a label, and the subtitle as the measure and unit. Returns the SVG and the chart spec. Charts: line, area, columns, bar, donut, treemap, waffle, funnel, pareto, scatter, heatmap, histogram, kpi, table, variance, bullet, waterfall, scorecard, combo, radar, candlestick, gauge, box, dotPlot.

Make a dashboard in one of the user's workspaces from queries on its datasets. type story: slides, each with a title and blocks, for a narrative (a monthly review). type board: one page of charts and KPIs with little text. Blocks: { kind: "chart", dataset_id, query (as aggregate takes it: group_by, grain, metrics, where, order_by, limit), chart: { chart, x, y, series, title, subtitle, format, mode, … as the chart tool } }, { kind: "kpi", dataset_id, query: { metrics }, title }, { kind: "text", text }, { kind: "commentary", prompt } (written from the numbers on its slide or board, on the user's own AI key at each refresh), { kind: "shape", shape: "rect" | "ellipse" | "line" | "arrow" }. Optional layout: story { x, y, w, h } in percent of a 16:9 slide; board { x, w } in 12 columns, { y, h } in rows. Leave it out and blocks are laid out for you. Optional style: { fontSize, titleSize, subtitleSize, color, fill, stroke, align, weight }. Charts are drawn at once; write each chart title as the takeaway. Returns the dashboard and its link; the user edits, comments and refreshes it in the app. Only the workspace's own datasets can be used.

Columns (id, original name, type, what each means, unit, role, whether it can be summed, range and examples), the team's approved metric definitions with their tie-out checks and glossary, links to other datasets (columns usable as alias.column), data quality (warnings, duplicate rows, empty values), freshness, and the server's current date for resolving relative dates.

Most frequent values of a column with counts. Use it to get exact spellings (e.g. "France" vs "FR") before filtering.

The dashboards in the user's workspaces: id, title, type (story or board), owner, when last refreshed, and the link.

Datasets this account can query, with row counts and when each was last refreshed.

Return matching rows. Filter with where, choose columns with select, sort with order_by. Columns of linked datasets work as alias.column. At most 1,000 rows; page with offset.

Re-run every query of a dashboard on the latest data, exactly as saved, and redraw its charts and numbers. Nothing is deleted: the data is untouched, and text blocks (commentary) keep exactly what they say; people rewrite commentary in the app. Returns what ran and the refreshed dashboard.

Tell the datatomcp team about a real problem with this data or these tools: a column with the wrong type, values missing that the user says exist, an error message you could not act on, or a result the user says is wrong. Use it rarely, only for problems you could not work around, never for ordinary questions or empty results. Describe what you called and what went wrong; do not include personal data or result values.

What you can answer, in one call: per workspace, every dataset with a one-line description, its columns (role, unit, whether they can be summed, what they mean) and freshness; the team's approved definitions; the links between datasets (columns usable as alias.column); the glossary; and five example questions, each with the exact tool call that answers it. Call it first.