The model can think and talk. Tools are how it reaches outside its own text box to actually do work, and they don’t make it smarter in the process.
Think of tools as arms and legs for the brain. A calculator doesn’t make a person wiser, it gives them a reliable way to calculate. A file reader doesn’t make an agent understand your business, it lets it read the files you gave it.
How it shows up
In Claude Code or Codex, tools are everywhere: reading a file, searching a codebase, running tests, opening a browser, calling QuickBooks through an API, or reaching Gmail through an MCP server. The moment the agent uses one, that’s a tool call, where it picks the tool, sends an input, gets a result, and decides what’s next. Connectors are how you attach a product like Gmail, Drive, or Slack, so the agent gets tools like search, read, create, and send.
Why you care
Tools have boundaries, and that’s the point. A tool might only read, or it might write, delete, or send messages, scoped to one folder or one channel. Once an AI has tools, your risk changes: it can change files, expose private data, or spend money through another service. That doesn’t make tools bad, it makes them serious. The practical question is always what the agent should be allowed to reach for: enough arms and legs to do the job, not every key in the building. Tools are the difference between an AI that comments on work and one that can actually do part of it.