Most interactions with large language models happen in a web browser. You paste code into a chat window, get a response, and paste it back into your editor. This workflow is useful, but it's disconnected from your actual development environment. The AI has no context for your project's structure, dependencies, or goals.

This shift is fundamental. A terminal-based agent can read your entire codebase, suggest changes across multiple files, run tests, and even commit code. It becomes an active collaborator rather than a passive code snippet generator. We'll be using Claude Code, a command-line interface (CLI) tool from Anthropic, to explore this new way of working.

Claude Code is distributed as an npm package, so you'll need Node.js installed on your system. Installation is a single command:

The -g flag installs it globally, making the claude command available anywhere in your terminal. Before you can use it, you need to authenticate with your Anthropic account.

This command will open a browser window for you to sign in. Once authorized, your credentials will be stored locally, and you won't need to log in again. You can verify the installation by asking for the version.

You can interact with Claude Code in two primary ways: interactively or non-interactively. Running the command by itself starts an interactive chat session right in your terminal.

This drops you into a chat prompt, similar to a web UI. Your conversation history is maintained for the duration of the session, giving the AI context for follow-up questions. To end the session, you can press Ctrl+C or type /quit. To clear the conversation history without ending the session, use the /clear command.

For quick, one-off tasks, non-interactive or 'print' mode is more efficient. Using the -p or --print flag, you can pass a prompt directly as an argument. The response is printed to standard output, and the program exits. This is ideal for scripting and automation.

The real power of a terminal-based agent comes from its ability to integrate with other command-line tools. You can use piping to send the contents of a file as context for your prompt. For example, to ask Claude to refactor a specific file:

Here, the cat command reads my_script.js, and the pipe character (|) sends that content directly to the claude command as input. The AI receives both the file's content and your instruction, all in one go.

For an AI agent to be truly useful, it needs permission to act. By default, Claude Code runs in a completely safe mode; it cannot read files, write files, or execute any shell commands without your explicit consent. When the agent wants to perform a sensitive action, it will pause and ask for your permission.

You'll be presented with several choices. You can approve or deny the action for this one instance (y/n), or you can grant or deny permission for all future actions of that type within the current session (a/v for always/never). This allows you to give the agent broad capabilities for a task you trust, or keep it on a tight leash for more sensitive work. To easily switch between these options in the terminal prompt, you can use Shift+Tab to cycle through the available choices before hitting Enter.