AI with Jspreadsheet Server

ULTIMATE

Jspreadsheet Server owns the live copy of every document, so its AI features work on the same spreadsheet the users are editing, under the same permissions, and their changes reach everyone in real time.

Three ways in

	AI agent	MCP server	PROMPT formula
Package	`@jspreadsheet/server-agent`	`@jspreadsheet/mcp`	`@jspreadsheet/openai`
Who talks to the model	Your server	The AI client	Your server
Where the user is	In your application, in a chat next to the spreadsheet	In Claude Code, Claude Desktop, an IDE or another agent	In a cell
What it does	Reads the document and edits it with tools, in a conversation	Exposes documents as MCP tools the client discovers and calls	Computes a cell from a prompt
Who pays for the model	You, with quotas through hooks	The client's owner	You

All three reach the document through the server, so they can be combined: a team works in the browser while an analyst's Claude Code edits the same workbook through MCP, and the in-app assistant answers questions about both.

What they share

The live document. No exports or stale copies: the AI reads what a person wrote a second ago.
Your authentication. Every request carries the user's token into the beforeConnect, beforeLoad and beforeChange hooks, so the AI can open and change exactly what its caller can. See Authentication.
One write path. An AI edit is an ordinary operation: validated, applied, persisted by your adapter and broadcast. Your storage never needs to know an AI was involved.
Formulas stay formulas. The agent and MCP write =SUM(B2:C2), not today's result, so the spreadsheet keeps working afterwards.

The AI agent

A streamed chat that turns plain-language requests ("add a total per region", "highlight the overdue tasks", "chart Q2 by product") into spreadsheet operations. It reads the document, calls tools until the job is done, and streams its answer. Requests can include attachments (Excel, CSV, PDF and images).

For products that expose AI to their own users, it provides the necessary controls:

allow refuses a prompt before it costs anything: per-user or per-address budgets, a site-wide budget, rate limits, attachment limits.
usage receives the token counts of every model call, for budgets and billing.
prompt keeps conversations in sessions, in your storage.

AI agent documentation

The MCP server

The mcp extension serves your documents over the Model Context Protocol at /mcp, with tools to list, read, create and edit them. Claude Code, Claude Desktop, IDE assistants and agents built on the MCP SDKs can use them with the token they carry.

claude mcp add --transport http sheets https://sheets.example.com/mcp \
  --header "Authorization: Bearer <token>"

MCP server documentation

The PROMPT formula

=PROMPT("Translate ", A2, " to French") joins its arguments into a prompt for an OpenAI model. The browser shows a placeholder while the server, which holds the API key, evaluates the prompt and broadcasts the result into the cell for everyone.

PROMPT formula documentation

Intrasheets

Intrasheets is the Excel-like application built on these pieces; its AI assistant is a chat pane that calls the agent. Try both on a shared workbook in the Intrasheets demo.

Cost and privacy

Budgets. The agent and the formula spend your model credits. Set quotas with the agent's allow and usage hooks before opening the assistant to the public. See Usage & billing.
Model choice. A small, fast model such as Claude Haiku covers most spreadsheet requests at a fraction of the cost; keep a larger model for heavy analysis.
Data. The model provider receives only the document of the request and its attachments. MCP clients receive only what their token can read.
Keys. Model keys live on the server, never in the browser.