MITHeadlessBuilt for Jev

Ask your data a plain question. See the view that answers it.

Get to the insight in complicated data, dense grids and tables. GridCue previews every change and never touches your data.

Try the full demo npm i gridcue Star on GitHub

12 rows

Account number	Household	Advisor	Registration type	Custodian	Market value	Concentration	Unrealized gain	Restricted holdings
AC-100000	Fairholm Household	Emerson Hale	Roth IRA	Northgate	$1,838,562	23.7%	-$56,460	No
AC-100001	Fairholm Household	Blake Moreno	Trust	Summit Trust	$3,757,780	13.8%	-$229,595	No
AC-100002	Hollis Household	Avery Lane	IRA	Summit Trust	$2,383,609	2.3%	$231,042	Yes
AC-100003	Alder Household	Avery Lane	Trust	Harborline	$402,764	1.2%	-$28,515	No
AC-100004	Jessup Household	Blake Moreno	IRA	Northgate	$1,227,557	2.3%	$101,445	No
AC-100005	Hollis Household	Blake Moreno	IRA	Summit Trust	$347,548	38.6%	$21,965	No
AC-100006	Dunmore Household	Avery Lane	Roth IRA	Summit Trust	$449,829	31.2%	$103,972	Yes
AC-100007	Fairholm Household	Emerson Hale	Taxable	Northgate	$4,458,649	0.7%	$18,629	No
AC-100008	Kestrel Household	Avery Lane	Roth IRA	Summit Trust	$1,903,520	12.5%	$271,507	No
AC-100009	Jessup Household	Avery Lane	Trust	Harborline	$1,390,109	41.2%	$46,702	No
AC-100010	Hollis Household	Blake Moreno	Roth IRA	Northgate	$612,040	20.2%	$120,725	No
AC-100011	Kestrel Household	Casey Ortiz	Roth IRA	Northgate	$4,940,891	4.5%	$685,245	No

Running in your browser on the keyless Mock Provider, over synthetic accounts. No server, no key.

How it works

From a sentence to a view you can trust

01
You ask
Type or dictate what you want to see: “Roth IRAs grouped by rep”. GridCue splits the request into parts and matches your app's own column names and values.
02
GridCue previews exactly what will change
The model only picks among choices your app declared. Code builds a checked plan and shows it line by line. Anything ambiguous becomes a question, never a guess.
03
You apply, and can undo
Nothing changes until you apply. The change is atomic, and one click puts the previous view back.

The difference

Why the model doesn't write the filter

	An LLM writes the grid state	GridCue
Who writes the grid state	The model generates it, and you hope the schema holds	Code writes it; the model only picks among closed choices
Before anything changes	Applied directly, often while it streams in	A deterministic Preview; apply only on approval, with undo
When it's unsure	It guesses	It asks one focused question, or declines

186 labelled live requests on a synthetic wealth schema, 0 wrong views with either strategy (jev-1.13.0, Sept 2026).

Safe by design

Built for grids where a wrong view costs something

View-only

Filters, sorts, groups and columns. It never edits a record.

Preview before apply

Every change is shown before it happens, and nothing applies without approval.

Your rows stay yours

Rows never leave your app. The model sees the request, the values read from it, and the columns you expose.

Undo, and no key in the browser

Every change can be undone, and your provider key stays on your server.

What GridCue is not: not SQL, not a chatbot. It doesn't edit your data, and it doesn't compute answers; it arranges the view that shows them.

Providers: Jev · Mock · your provider

Grids: TanStack Table · any array · AG Grid (next)

Fits your domain

Your words, your categories, a few lines of setup

Every app says things its own way. Three optional declarations teach GridCue your domain: what a row is, which columns name other records, and your own categories over a column's values.

Choose how GridCue asks Jev: “fan-out” handles compound requests, nesting and “also”; “focused” asks fewer questions for simpler grids.

defineSchema(columns, {
  rowNoun: "account",                    // "biggest accounts first" means the rows
  columns: {
    household: { entity: "household" },  // "largest households first": asks, never guesses
    registration_type: {
      valueGroups: [{ label: "Retirement", values: ["ira", "roth_ira"] }],
    },
  },
});

createJevProvider({ apiKey, strategy: "fan-out" }); // or "focused": fewer questions

Quick start

A working command bar in five minutes, no key

npm i gridcue

import { createGridCue } from "gridcue";
import { createMockProvider } from "gridcue/mock";
import { GridCueBar } from "gridcue/react";
import { createTanStackAdapter, gridcueFilterFn, schemaFromTanStack } from "gridcue/tanstack-table";
import "gridcue/styles.css";

const table = useTable({ features, columns, data, defaultColumn: { filterFn: gridcueFilterFn } });
const [cue] = useState(() => {
  const schema = schemaFromTanStack(table);
  return createGridCue({ schema, adapter: createTanStackAdapter({ schema, table }), provider: createMockProvider() });
});

<GridCueBar controller={cue} />

The Mock Provider runs in the browser. When you're ready, turn on Jev behind your own endpoint. Read the docs.