Tools, connectors, and the outside world

A prompt-only chat is like an intern locked in a room with no phone. Powerful, but blind. Tools are how you unlock the door.

What "tools" mean in AI

A tool is a function the model can decide to call. You describe the function in JSON — its name, what it does, and its parameters — and the model, mid-response, chooses to call it. The runtime executes the function, hands the result back, and the model continues.

Defining a tool

JavaScriptconst tools = [
  {
    type: "function",
    function: {
      name: "get_weather",
      description: "Get the current weather for a US city.",
      parameters: {
        type: "object",
        properties: {
          city: { type: "string", description: "City and state, e.g. 'Denver, CO'" }
        },
        required: ["city"]
      }
    }
  }
];

Two rules matter more than anything else:

1. Descriptions are prompts too. The model reads your tool descriptions when deciding when to call. Write them like documentation for a smart intern.
2. Fewer tools per turn. Give the model 5 great tools, not 50 mediocre ones. Group functionality: search_files beats search_gdrive + search_dropbox + search_notion.

Connectors: pre-built tools for popular apps

You could build a Gmail tool from scratch. Or you could use a "connector" — a pre-built package of tools that already handles auth, pagination, and rate limits.

Browsing the web (autonomously)

The other superpower is browser use. Some tools give the model a real Chromium browser. It can click, type, scroll, screenshot, and log in. This is what Perplexity Computer, Claude Computer Use, and OpenAI's browser tool do.

Rule: use browsers only when APIs don't exist. APIs are 100× cheaper and 100× more reliable. Reach for browser use only for sites without APIs (many state RFP portals, some analytics tools).