open-sheet

open-sheet is an open-source TypeScript framework that lets AI coding agents generate spreadsheets without ever writing a cell address. Instead of producing brittle =SUM(B2:B13) formulas that break when rows shift, agents describe the model in code and open-sheet resolves all coordinates at compile time, exporting a live .xlsx with real formulas, conditional formatting, defined names, and print settings. The project is MIT-licensed and published on npm under @open-sheet/core, @open-sheet/cli, and @open-sheet/mcp.

What It Is

open-sheet is a spreadsheet compiler for agent pipelines. You write column definitions and data in TypeScript/React — using references like r.cell('revenue') and r.prev().cell('revenue') instead of A1 coordinates — and the framework places every block on the grid, resolves every reference, and emits a .xlsx that recalculates in Excel, Google Sheets, and LibreOffice. The core insight is that cell addresses are the one thing an agent cannot reliably hold onto: insert a row and every hard-coded reference in the file becomes silently wrong. open-sheet removes that failure mode by owning all coordinates itself.

How the Address-Free Model Works

The framework's central abstraction is a reference API that never exposes grid coordinates to the author:

• r.cell('revenue') — the revenue cell in the current row
• r.prev().cell('revenue') — the same column, one row up
• ref('pl').column('revenue') — the full revenue column from another block
• ref('assumptions').get('growth') — a scalar assumption emitted as an Excel defined name

These resolve to A1 only after layout has placed every block. Scalar assumptions become Excel defined names in the exported file, so a recipient opening the workbook reads =B4*growth rather than =B4*B2. Adding a row to the data array causes every reference to re-resolve automatically — no formula editing required.

Architecture: One Formula Tree, Two Backends

Every formula is an expression AST with two consumers: serialize() writes an Excel formula string into the .xlsx, and evaluate() computes the number shown in the browser viewer. The project validates both backends in CI by exporting a workbook without cached results, recalculating it in LibreOffice, and diffing the output against the framework's own evaluator. The homepage states that 127 functions are whitelisted and that a disagreement fails the build. When the evaluator cannot compute a cell it emits #NOT_EVALUATED rather than a plausible-looking number, making failures visible rather than silent.

Agent-Native Workflow

The scaffolder (npx @open-sheet/cli init my-sheets) creates a workspace with five agent skills covering authoring, block placement, formatting, recipient capabilities, and print setup. An MCP server (@open-sheet/mcp) exposes the same operations over Streamable HTTP so any agent framework can drive the dev server programmatically; stale writes are refused with a 409 rather than silently overwriting work. Inspect mode lets a reviewer click any cell in the browser viewer to see its source line, resolved formula, and computed value, and leave comments for the agent to apply.

Update: v0.1.8 — Six Rounds of Real Workbooks

The latest release is @open-sheet/core@0.1.8, published on 2026-08-20 and described as "six rounds of real workbooks." The homepage ships six dogfood workbooks — a departmental budget variance report, a SaaS KPI dashboard, a fixed asset register, an FY26 budget, and two Taiwanese business documents — totalling 171 live formulas across six files, none of which contain a hand-authored cell address. The GitHub README notes the project is in early development and directs users to follow milestones; the npm packages are live and the npx @open-sheet/cli init path is the recommended entry point. The project follows in the lineage of open-slide (presentations) and open-doc (documents) as the third medium in that agent-document framework family.