Getting started with Agent Tools

agent-tools gives an AI agent the ability to work with spreadsheets just like a human analyst. Read, calculate, write and hand back the updated spreadsheet for a human to open in Excel. It runs headless and in-process — no Excel install, no COM, no cloud round-trips.

There are two ways to use it: run it as an MCP server, or import the tools and wire them into your own agent SDK.

Install

npm install @grid-is/agent-tools

executeOfficeJs needs the optional @grid-is/apiary-officejs; the other tools don’t.

Run as an MCP server

Start it straight from npx:

npx @grid-is/agent-tools

Or point any MCP client at it:

{
  "mcpServers": {
    "grid": {
      "command": "npx",
      "args": ["-y", "@grid-is/agent-tools"]
    }
  }
}

The server speaks MCP over stdio, so it works with any MCP-capable runtime — Claude, the OpenAI Agents SDK, LangChain, or your own product.

Wire the tools into your own SDK

Every tool is a plain object — { name, description, inputSchema, run } — so you can map them onto any framework’s tool interface. createGridTools() returns the full set, statefully tracking the loaded workbooks:

import { createGridTools } from "@grid-is/agent-tools";

const tools = createGridTools();
const loadWorkbook = tools.find((t) => t.name === "loadWorkbook");
const result = await loadWorkbook.run({ path: "budget.xlsx" });
console.info(result.text);

inputSchema is a zod raw shape, so z.toJSONSchema(z.object(tool.inputSchema)) produces the JSON Schema most SDKs expect for a tool’s parameters.

serveStdio() and toMcpServer() are exported too, so you can register a subset of tools, rename the server, or add your own:

import { serveStdio, createGridTools } from "@grid-is/agent-tools";

await serveStdio({
  name: "my-agent",
  tools: createGridTools().filter((t) => t.name !== "executeOfficeJs"),
});

Where to next

Tools — the full catalogue, grouped by what an agent uses them for.
API reference — every exported function and type.