A spreadsheet your AI agent can drive
VisiGrid ships a first-party MCP server. Claude Code, Claude Desktop, Codex CLI, and any local MCP host read and edit a live native window — cells change on screen as the agent works, and every edit is one undo away.
claude mcp add visigrid -- vgrid mcp
That's the entire setup for Claude Code. Codex: codex mcp add visigrid --command vgrid --args mcp. Linux, macOS, Windows.
Three steps, no tokens
Add the server
One line registers VisiGrid with Claude Code. No tokens, no config files, no API keys.
claude mcp add visigrid -- vgrid mcp Ask for what you want
"Clean up column C." "Build a 12-month projection." The agent reads and writes your open workbook.
Click Allow — once
The first time an agent touches your spreadsheet, VisiGrid asks. Approval persists until you revoke it.
vgrid pair --revoke "Claude Code" What a session looks like
The agent orients with a read, batches its edits, verifies its own work — and the whole time, you're watching it happen in a real window.
Agent access you can trust
An agent editing your spreadsheet should be a feature, not a leap of faith. Four guarantees, enforced by the engine — not by prompt engineering.
You can see it
Agent edits render live in the window you’re looking at. There is no headless back door to an open workbook.
You can undo it
Every batch of agent edits is a single entry in your undo history. One Ctrl+Z and it never happened.
It can’t clobber you
Writes carry a revision check. If you edited the sheet since the agent last read it, the write fails instead of overwriting your work.
It can’t sneak in
Access is consent-gated per client and revocable at any time. Until you click Allow, the socket answers nothing.
Built in, not bolted on
Excel MCP servers exist — as third-party bridges: headless file manipulators, or Windows-only COM automation. VisiGrid's server ships inside the app, works everywhere the app does, and is wired into its consent, undo, and concurrency machinery. Local-first, so your rows never transit anyone's cloud.