The agent-first
spreadsheet engine.
Your agents, your spreadsheets, your apps. Together.
Businesses run on spreadsheets.
Your apps and your agents work in every system. They can't run your spreadsheets.
The app shows a grid. The agent reads a file. You import, export and copy-paste.
Agents get full context. Humans keep full control.
Humans plan. Agents do. Spreadsheets show the work.
Because the agent runs the model, not the file.
An open spreadsheet changes three things about how your agent works.
Same model. Different surface.
A public benchmark, the same tasks and the same models. The gap is the surface the agent works through, not the model.
How the run was madeSpreadsheetBench · verified-400
tasks solved · October 2026- SheetXL97%
- Grid Agent91.25%
- Shortcut.ai86%
- Claude Opus 4.6, bash agent28.75%
- Claude in Excel15.36%
- GPT-5.2, bash agent14.29%
SheetXL: our SpreadsheetBench submission, November 2026. The highest score of any vendor that has published its results. Other figures from the public leaderboard (V1 Verified), which also carries higher entries with no published outputs or method.
- SheetXL3.3k
- Agent alone≈23.1k
Also fewer tool calls and faster per task, for the same reason.
Agent alone: Claude Code reading the file with Python, no engine. SheetXL measured per task over the same 400 SpreadsheetBench tasks, October 2026. The agent-alone figure is derived from Grid Agent's published savings over a bare Python agent and SheetXL's measured 3.0× over Grid Agent.
Why not a standalone agent, or an Excel plugin?
One never gets into the spreadsheet. The other never gets out of it.
| Claim | SheetXL | Standalone agent | Excel plugin |
|---|---|---|---|
| The agent sees all of your work: your systems, your files and the live spreadsheet | |||
| You and the agent work in the same grid, live | |||
| An audit log of every change the agent made, and why | |||
| Undo any agent change, or roll back the whole session | |||
| Every change recalculated and verified by a real engine | |||
| Works with the agent you already run | |||
| Runs inside your own app | |||
| Runs on your server or air-gapped, and the file never leaves | |||
| Open UI and skills you can read, in the build that scored the benchmark |
Standalone agent: Claude Code or Codex reading the file with Python. Excel plugin: Copilot, Claude in Excel.
Copilot and Claude in Excel are an agent plugged into a spreadsheet. SheetXL is a spreadsheet plugged into your agent.
It is a surface problem, not a model problem. The best model in the world working through Excel's surface scores 15.4%. The same model on SheetXL scores 97%. Better models do not change the surface.
Three ways in.
Add it where your agents, your apps or your servers already are. Same engine underneath, one file in and out.
Your agents. Your spreadsheets. Your terms.
One prompt installs it. Then ask your agent something you would normally check by hand.