Best way to build
spreadsheets with AI.

The .xlsx file itself stays on disk — no GridPath cloud, no upload, no third-party server in the middle. What does flow to the LLM provider you've signed in with (Anthropic for Claude, OpenAI for ChatGPT) is your prompt, sheet structure, and whatever cells the agent reads or writes during the run. Same shape and same privacy terms as using that provider's product directly. See Privacy for the long version.

GridPath connects to your Claude or ChatGPT subscription — you pay the LLM provider whatever you already pay them, and nothing extra to us.

Yes — but with one honest caveat. Saves are surgical: a native patcher rewrites only the cells and structure the agent actually changed inside the original .xlsx package. That means everything else — charts, conditional formatting rules, named ranges, data validation, merged cells, drawings, freeze panes, pivot tables, themes — comes out byte-identical. Open the file in Excel after and it looks exactly like it did before, plus your new edits. Every save is also checked against Excel's strict format rules before it's written — if the output wouldn't open cleanly, GridPath refuses to write it. And for the rare edit the patcher can't represent, GridPath never silently overwrites a file with content at risk — it names exactly what a reduced-fidelity copy would lose and leaves the original untouched. More on how surgical saves work.

The caveat: GridPath's in-app grid (built on Univer) doesn't visually render charts, conditional-formatting coloring, drawings, or pivot tables while you're editing. They're fully intact in the file — just not painted on screen inside our app. The agent still respects merged cells, freeze panes, and named ranges natively. If your workflow needs to see chart updates live, do that pass in Excel; if it's structural / formula work, GridPath is faster.

Three ways. (1) Reject the proposed batch before it commits — nothing touches the file. (2) ⌘Z to undo accepted changes within the session. (3) After save, we keep a .bak with the previous state for one-step recovery, even after reopening.

You can, and for some tasks it works. The problem is what you can't see. A general-purpose agent edits your workbook by loading it through a spreadsheet library, mutating it, and writing the whole file back out — and anything the library doesn't model (plugin data, printer settings, and in real files much more) silently vanishes on the way through. The agent can't warn you, because its tools never showed it that content existed. We benchmarked this with the same Claude model on both sides, twice: on our test workbook the CLI agent dropped file content in 5 of 5 runs on the older model generation and 2 of 5 on the newest — reporting success every time — yet on a big intimidating bank model it chose a careful manual approach and preserved everything. Its safety is an invisible per-file judgment call that got luckier. GridPath saves through a surgical patcher instead — everything you didn't touch is copied byte-for-byte, every run, on every file, by architecture rather than by judgment.

Excel Copilot is a single-turn assistant inside Excel. GridPath runs a multi-turn agent loop — fetch web data, write 58 rows, apply 64 formatting ops, fix its own formula errors — across 10–15 tool calls per prompt, until the model is genuinely done. Plus you pick the underlying LLM and review every edit before it lands.

.xlsx today. .xls, .csv, and Google Sheets via API are on the roadmap.

Yes, two ways. Open several workbooks in tabs — each gets its own agent session, running in parallel. And within a session, attach up to 5 other .xlsx files as read-only references: the agent sees a compact preview of each and reads exact ranges on demand, so it can compare your model against an analyst's, pull assumptions from last quarter's file, or cross-check comps — without you copy-pasting anything between windows.

Reference files are never modified, and only the cells the agent actually reads go to the model — which keeps prompts small and turns fast, even with several large workbooks attached.

Yes. Cross-sheet references (=Sheet2!A1) work natively, and renaming a sheet auto-updates references in other sheets. External workbook links (=[Other.xlsx]Sheet1!A1) survive as cached values but don't recompute live — same limit Google Sheets has.

macOS (Apple Silicon, macOS 12+) and Windows (10/11). Linux is on the roadmap. Source available for security review on request.