← All posts

Why we built GridPath.

THE GRIDPATH TEAM · JULY 4, 2026

We live in spreadsheets, and we wanted AI that actually works in them. Not a chatbot you paste cells into — an agent with real control over its context and tools built for the grid. Here's the thinking behind it.

The problem: spreadsheets aren't walls of text

Every general-purpose AI tool handles a workbook the same way: flatten it into text, stuff as much as fits into the prompt, and hope. That fails in two directions at once. A real workbook — fifty sheets, tens of thousands of formulas — blows straight past any context window, so the tool silently truncates and the model reasons about a spreadsheet it has only partially seen. And even when the file does fit, you're paying for tens of thousands of tokens the model doesn't need, on every single turn, mostly to re-read cells that never changed.

The output side is just as bad. A chat assistant "edits" a spreadsheet by emitting cell contents one message at a time, usually as pasted values that go stale the moment an input changes. Excel Copilot is closer, but it's single-turn: one instruction, one edit, no loop.

We built GridPath around two convictions: you should control exactly what context the model sees, and the agent should work through tools specialized for spreadsheets. Everything else follows from those two.

Context is something you manage, not something you dump

GridPath treats the model's context window as a budget and spends it deliberately:

• Compact previews, not full dumps. Each turn ships the workbook's structure and a sparse preview of what's actually in it — non-empty cells, formulas alongside their computed values — instead of a serialized copy of every sheet.
• Selection as context. Highlight D3:D8 in the grid and it lands in the conversation as a chip. "Fix these" means those six cells — no describing ranges in prose, no pasting.
• Reference workbooks. Attach up to five other .xlsx files to a session as read-only context — an analyst model to compare against, last quarter's actuals, a comps sheet. The agent gets a tight preview of each and pulls exact ranges on demand with a read_reference tool. Only the cells it actually reads ever reach the model, and the files themselves are never touched.
• On-demand reads for everything else. When the agent needs detail beyond the preview, it asks for a specific range. The model earns its context instead of drowning in it.

This is the same lesson coding agents learned: don't paste the whole repository into the prompt — give the agent a map and let it open files. GridPath applies it to workbooks.

Tools shaped like spreadsheet work

A spreadsheet agent that writes one cell per tool call is a toy. Real work is bulk: build a 58-row income statement, apply number formats to a whole block, restyle every header. GridPath's tools match that shape — one set_range call writes a 58×9 block, one set_format call applies 64 formatting operations. What would be sixty round-trips in a generic tool-use setup is one.

The tools are also formula-first. The agent writes =SUM and =VLOOKUP, not pasted numbers, so what it builds stays live — change an assumption and the waterfall recalculates. And because the loop is multi-turn, the agent fetches web data, writes the block, formats it, notices a broken formula, and fixes it — typically 10–15 tool calls per prompt, until it's genuinely done.

What that buys you

More detailed work. Token budget the model doesn't burn re-reading your workbook is budget it spends on the task — deeper analysis, more scenarios, better labels, formats you'd actually ship.

Faster turns. Batched tools mean fewer round-trips. Stable, compact context means ~75% of each prompt hits the provider's cache. The app itself is a native Rust core, not Electron, so nothing between you and the model is the bottleneck.

Token efficiency that's actually your money. GridPath runs on your own Claude or ChatGPT subscription — we don't resell tokens or add a margin. So every token saved is your quota, not our revenue. That alignment is deliberate: it keeps us honest about efficiency in a way "AI credits" pricing never would.

And you stay in control

Agents earn trust by being reversible. Every edit lands as a diff-first proposal — green pending cells you accept or reject before anything commits. ⌘Z undoes accepted changes, and a .bak of the previous state survives even after you save and reopen. Your .xlsx never leaves your disk; what goes to the model is the prompt, the structure, and the cells the agent reads or writes.