Get started with Claude Code and Codex in the terminal

The useful step from chatting about code to working agentically is giving the assistant a real project and a way to check its work. It can then read files, propose changes, run commands and respond to the results. That makes the working directory and the brief rather important.

This is the setup available in October 2025. It is a historical starting point, so use the current vendor installation pages if you are setting up a machine today. In particular, an unpinned npm command run today installs today's package, not the version available when this guide is set.

Start with one local project

Use a development copy of a repository you are allowed to work on. Open a terminal in that directory and check Git before launching either assistant. My workspace contains separate sites and shared plugins; starting one directory too high can give the agent the wrong idea about what belongs together.

cd /path/to/your/project
git status --short
node --version
npm --version

The commands below use npm, the package manager supplied with Node.js. The Claude Code README at the time specified Node.js 18 or later. Use a compatible maintained Node release for your machine, and follow the historical platform requirements if reproducing an old environment. These examples assume a macOS or Linux shell; Windows users should consult the platform-specific setup rather than assume every shell behaves the same way.

Install and open each CLI

Both vendors documented these package names in October 2025:

npm install -g @anthropic-ai/claude-code
claude --version
claude

Complete Claude Code's account sign-in when prompted. In a separate terminal, or after exiting that session, install Codex:

npm install -g @openai/codex
codex --version
codex

Codex's October README offered sign-in with ChatGPT for eligible subscriptions, with API-key authentication as a separate setup. These are coding clients, rather than an installation of the ordinary ChatGPT website. Check the account and billing route you select; access to one product does not establish access to every API.

Make the first task an inspection

Read this repository and its instructions. Tell me how to run it locally, where the relevant code lives and what checks are available. Report existing changes. Do not edit files yet.

That answer gives you something concrete to correct before the assistant starts. If it identifies the wrong application or proposes a test command that does not exist, resolve that first. If a command is missing after installation, check the installation output and PATH in a new terminal instead of repeatedly installing with elevated permissions.

Choose a change with an obvious finish line

For the first edit, pick something like a misleading validation message and explain the expected result. Ask the agent to make the change, run the relevant check and show the diff. Review it with git diff and try the affected screen yourself. Keep deployment as an explicit next step.

I would get comfortable with one agent on one small task before asking two tools to touch the same project. Once the setup works, give both agents the project's working rules.