The verified Excel specialist.
A local, tool-calling AI specialist that lives as a chat pane inside Excel on Windows.
Design-partner programme. This product is new, and we say so. We're onboarding a small number of design partners: pilots scoped with our engineers, early-adopter pricing, and a roadmap shaped by your feedback.
Apply as a design partnerHow it works
Ask in plain language
Type a request in the chat pane inside Excel. It works in any language — Hinglish verified.
It plans the tool calls
Vexilo inspects and reads the sheet, then chooses from 19 tools — write/formula, sort, format/chart, sandboxed Python, and more.
It runs the result in Excel to prove it
Vexilo uses Excel's own engine to evaluate and recalculate — it doesn't just print an answer.
Review the diff
Every cell it touched is shown in a diff before you keep it.
Accept or undo
One Ctrl+Z undoes the whole reply. Nothing is saved to your open file automatically.
Reliability you can verify, not just trust.
Runs on your machine
A local model (27B, Apache-2.0, via llama.cpp) — no account, no sync, nothing leaves your network.
Proves its own work
It runs every result in Excel to prove it, and shows a diff of every cell touched.
Safe to hand off
Additive by default, asks before destructive actions, ignores prompt injection in cells, and one Ctrl+Z undoes a whole reply.
Everything Vexilo does
19 tools: inspect/read, write/formula, sheets/sort, format/chart.
Sandboxed Python for anything the built-in tools don’t cover.
Excel-engine evaluate/recalculate — the sheet itself proves the answer.
Ask-user and undo are first-class tools, not afterthoughts.
Additive by default; asks before destructive actions.
Ignores prompt injection found in cells.
Never saves your open file automatically.
One Ctrl+Z undoes a whole reply.
Runs on your machine on a local model (27B, Apache-2.0, via llama.cpp).
No account, no sync — nothing leaves your network.
Tested against a real persona task set with zero silent wrong answers.
Works in any language — Hinglish verified.
How Vexilo keeps its answers honest.
Grounded in your documents, checked by deterministic rules, held for a human when unsure, and recorded — the model's own output is never the last word.
- Vexilo runs the result in Excel's own engine to prove it — it doesn't just print an answer.
- Every reply shows a cell-level diff of exactly what changed, before you keep it.
- One Ctrl+Z undoes a whole reply, and it asks before destructive actions.
Human-in-the-loop, by default.
Every Mobiprise product is designed to be accountable to the humans who run it — not just fast.
- Additive by default, and asks before destructive actions.
- Ignores prompt injection found in worksheet cells.
- One Ctrl+Z undoes a whole reply; it never saves your open file automatically.
- Runs on a local model (27B, Apache-2.0, via llama.cpp) — no account, no sync, nothing leaves your network.
What it runs on
- Windows, Excel 2019+ / 365
- An NVIDIA GPU is strongly recommended — the difference between seconds and minutes
Built for these teams
- Finance and spreadsheet teams doing formula- and data-heavy work in Excel.
- Anyone who needs a change proven inside the sheet, not just described in a chat reply.
- Non-English-speaking users — works in any language, Hinglish verified.
Questions worth asking
It's in early access: in development, with substantial R&D ground already covered. We're open to pilots and collaboration, with general release coming soon.
No. It runs on your machine on a local model (27B, Apache-2.0, via llama.cpp). No account, no sync — nothing leaves your network.
It's additive by default, asks before destructive actions, and ignores prompt injection found in cells. One Ctrl+Z undoes a whole reply, and it never saves your open file.
Every reply that touches cells shows a diff of exactly what changed before you keep it.
It has 19 tools: inspect/read, write/formula, sheets/sort, format/chart, a sandboxed Python environment, Excel-engine evaluate/recalculate, asking you a question, and undo.
Windows, Excel 2019+ or 365. An NVIDIA GPU is strongly recommended — it’s the difference between seconds and minutes. We recommend running a 27B LLM, which runs comfortably on a machine with 48 GB of VRAM — smaller models also work, with some compromise on speed and accuracy.
Join Vexilo's early access.
We're open to collaboration now, with general release coming soon.