A first coding agent session on your own computer

Your first coding agent session can begin with a small task in an ordinary project folder. You might ask it to explain a script, fix a broken button or add a missing check. On Windows, macOS and Linux, the basic routine is similar: open the project, describe the task, let the agent inspect the relevant files and review its work.

The setup differs between operating systems, but you do not need to understand every terminal command before starting. A little preparation makes the first session easier to follow.

Understand what the agent can do. A coding agent connects an AI model to tools on your computer. Depending on its permissions, it can read files, suggest or apply edits, run commands and check the results. Claude Code, Codex and Gemini CLI are examples of tools built around this kind of workflow.

The model usually runs on a remote service. The agent runs locally and sends relevant context to that service. This can include source code and command output, so start with a project you are comfortable sharing under your provider’s terms. Keep passwords, private keys and unrelated personal files out of its working folder.

Choose a working environment. Follow the current installation instructions for your chosen agent. Supported operating systems, required software and sign-in methods can change. These are the main differences to keep in mind:

• Windows: Check whether the agent supports PowerShell directly or recommends Windows Subsystem for Linux, usually called WSL. WSL provides a Linux environment inside Windows. If you use it, install and run the agent within that environment, and follow its guidance on where to store projects.
• macOS: You will usually work in Terminal or another terminal application. Some projects also need Apple’s command line developer tools. Install additional dependencies when the project or agent documentation calls for them.
• Linux: Use your distribution’s terminal and package manager for prerequisites. Installation commands differ between distributions, so check instructions before copying them. Avoid running the agent with administrator privileges unless a specific task requires them.

A terminal is simply a text interface for running programs. Before launching the agent, navigate to your project folder. That location helps establish which files the agent should work with, although its actual access depends on the tool’s permissions and sandbox settings.

Connect the tool to a model. Some agents use an account subscription; others accept an API key and a configurable service address. An API key is a credential that authorizes requests and can incur charges. Store it using the tool’s recommended method, and keep it out of source files and version control.

Vana API provides one endpoint at https://api.vanaapi.com with support for the OpenAI-compatible API, Anthropic Messages and Gemini native protocols. It works with Claude Code, Codex, Gemini CLI and applications built with an OpenAI SDK. The configuration fields vary by tool, so use the relevant Vana API setup instructions for the address, key and model identifier.

This gives you a way to access models such as Claude Sonnet 5.5 through a shared service. Vana API bills for actual token usage with no monthly fee. Agent sessions can make several requests while inspecting files and checking changes, so a short instruction can still produce substantial usage. Create a Vana API account when you are ready to connect your chosen tool.

Give the first session a narrow task. Try: “Explain how this project starts. Read the relevant files, but do not change anything.” Then ask for one small edit with a clear result. For example: “Show a helpful message when the input is empty, and run the existing tests.”

Read permission prompts before accepting commands. After an edit, inspect the changed files and ask what checks ran. If the agent could not run a test, that limitation matters. Save a known working version with Git or a backup before continuing. Small tasks make the results easier to understand and give you a repeatable routine on any of the three operating systems.