Describe the model you need in natural language — your coding agent writes the React. open-sheet owns the cell addressing, the formula references, the recalculation and the export. No A1 address is ever authored by hand.
The reason is specific: =SUM(B2:B13). Cell addresses are the one thing an agent cannot hold onto. It miscounts the header row. It forgets the total moved when the data grew. It writes a reference to B7 and then inserts a quarter above it — and now every formula in the file is quietly wrong. Not crashed. Wrong, which is worse.
So open-sheet takes the addresses away.
col('quarter', { header: 'Quarter' }),
col('revenue', { header: 'Revenue', format: 'currency', bar: true }),
col('cogs', { header: 'COGS', format: 'currency' }),
col('grossProfit', {
header: 'Gross profit',
format: 'currency',
formula: (r) => sub(r.cell('revenue'), r.cell('cogs')),
}),
col('margin', {
header: 'Margin',
format: 'percent',
scale: ['#fee2e2', '#ffffff', '#dcfce7'],
formula: (r) => div(sub(r.cell('revenue'), r.cell('cogs')), r.cell('revenue')),
}),
col('qoq', {
header: 'QoQ growth',
format: 'percent',
highlight: { above: 0.15, fill: '#dcfce7', bold: true },
formula: (r) => (r.isFirst ? null : sub(div(r.cell('revenue'), r.prev().cell('revenue')), 1)),
}),
| Quarter | Revenue | COGS | Gross profit | Margin | QoQ growth |
| Q1 | 12,400,000 | 5,100,000 | 7,300,000 | 58.9% | |
| Q2 | 13,900,000 | 5,600,000 | 8,300,000 | 59.7% | 12.1% |
| Q3 | 15,200,000 | 6,050,000 | 9,150,000 | 60.2% | 9.4% |
| Q4 | 18,650,000 | 7,100,000 | 11,550,000 | 61.9% | 22.7% |
| Total | 60,150,000 | 23,850,000 | 36,300,000 |
Add a quarter to the data array and every reference re-resolves at compile time. There is no B2:B13 to get wrong.
Gross profit is =B2-C2 in the file, not 7,300,000. Change a number and the whole sheet recalculates — in Excel, in Sheets, in LibreOffice.
Different tool, different job. Claude for Excel makes one person faster at making a spreadsheet. open-sheet makes the spreadsheet not need making a second time.
| Claude for Excel | open-sheet | |
|---|---|---|
| Input | a workbook that already exists | a source file |
| Who is in the loop | a person, every time | a person once, at review |
| Reviewing a change | a binary diff | a code review |
| Doing it 500 times | 500 sessions | a loop |
| Needs | Excel, an account, a paid plan | Node. MIT. |
An assistant working inside a spreadsheet still writes =SUM(B2:B13) into a cell. That is a hand-authored address — authored by a model, but hand-authored. Insert a row and it is wrong, and nothing tells you. This is not a question of the model being clever enough: A1 is the only language that medium has.
And when open-sheet cannot compute a cell, it says so. #NOT_EVALUATED, never a plausible-looking number. Anything writing into a live cell produces something number-shaped whether or not it was sure — which is the failure that costs the most and shows the least.
Where Claude for Excel wins, and it is not close. It writes live formulas too — the difference is never “ours recalculate and theirs do not.” A workbook you did not generate, we cannot read at all. It computes with Excel’s own engine, so its numbers are Excel’s numbers by construction, where we have to earn that with a cross-engine check on every build. Five hundred-odd functions to our hundred-odd. Pivot tables.
Use that when you have a spreadsheet and want help with it. Use this when the spreadsheet is the output of a pipeline — a monthly board pack, five hundred invoices, a model regenerated whenever the data lands. The two compose, in that order.
The .xlsx holds formulas, not baked numbers. Change an assumption and the sheet recalculates in whatever they opened it with.
validate puts
a fixed set of answers on a column, with both the prompt and the refusal message.
protect unlocks the
cells they are meant to fill in and leaves every formula shut.
appendable emits an Excel Table,
so a row typed below the last one is taken into every range that reads it.
note attaches provenance
to the cell itself — shown on hover, not buried in a footnote.
Headers, footers, page numbers and page breaks that name blocks rather than row numbers, so they move when the content does.
npx @open-sheet/cli init my-sheets writes a workspace
carrying five skills — how to author, how to place blocks, how formats work, what the
recipient can do with the file, and how it prints.
/create-sheet and describe the model. The skill opens by
confirming the material and refusing to invent numbers, which is the failure this
medium punishes hardest.
Click any cell in the viewer to see its source line, its resolved formula and its computed value — or leave a comment for the agent to apply.
An MCP server over Streamable HTTP exposes the same operations the browser uses, so a stale write is refused with a 409 rather than silently overwriting somebody.
Our evaluator and Excel have to agree about what a formula means, or the number in the viewer is a guess. Every whitelisted function is compiled into a fixture, exported without cached results, recalculated by LibreOffice, and diffed against what we computed. A function joins the whitelist only when a second engine agrees.
That check is also why SORT
is implemented here rather than taken from a library: the library’s compares as
strings, which is indistinguishable from correct until your numbers stop having the same
number of digits.