Codex CLI Computer Use: Availability, Setup, and Limits
Codex CLI can use computer-use tools to interact with a browser or desktop application. This guide explains how to verify what is supported, configure a bounded task, and keep human decisions in the loop.
Updated 6 min read
Codex CLI computer use means giving the agent access to tools that can interact with a browser or a graphical desktop application: clicking, typing, reading the screen, and navigating UI elements. Whether a particular computer-use capability is available depends on your platform, the model you are using, and the version of Codex CLI you have installed.
Before you enable any computer-use tool, decide exactly what actions you are willing to let the agent take and record that decision somewhere the next session can find it. A tool like fenbs keeps approved action limits as Decisions and rules, so a future session does not have to ask you to set the boundaries again.
What computer use adds to Codex CLI
Standard Codex CLI tasks involve reading and writing files and running terminal commands. Computer use extends that to visual interfaces: opening a URL, filling a form, clicking a button, or reading information from a web page that does not have an API.
This is useful for tasks that live in a browser: filling in a form, navigating a tool that has no CLI, or confirming that a deployed change looks correct. It is not a replacement for scripted automation when an API or a terminal command is available; the computer-use path is slower and less reliable for tasks that can be done programmatically.
Verify what is supported on your setup
Computer-use support in Codex CLI depends on the model in use and the platform it is running on. Not every model supports computer use, and local sandboxing requirements vary by operating system. Check the OpenAI Codex documentation (opens in a new tab) and your API account to confirm which capabilities are available.
Run codex --version to confirm which version of the CLI you have. If computer use is not listed in the tools available to the agent, update the CLI first: npm install -g @openai/codex to get the latest version.
On macOS, Codex CLI uses Apple's sandbox framework for local execution. On Linux it uses Docker or Podman. Computer-use tools that require a graphical environment need a display server to be running or available in the sandbox. Check the requirements for your specific setup before assuming computer use will work out of the box.
Configure a bounded computer-use task
Start with the narrowest task you can define. Computer-use sessions are harder to review than file-edit sessions because the actions happen in a visual interface rather than in a diff. The more specific the goal, the easier it is to check that the agent did exactly what you intended.
A good bounded task looks like: "Open this URL, fill the Name field with X, fill the Description field with Y, submit the form, and confirm the success message appears." A poor bounded task looks like: "Check the app and fix anything that looks wrong."
Use approval mode to review each action before it happens. In suggest or plan mode, the agent shows you what it intends to do before it acts. For computer-use tasks, this is especially important because some actions, like submitting a form or clicking a delete button, cannot be undone from the CLI.
Approve limits before a session starts
Decide before the session what the agent may and may not do, and write it down. Examples: "May open any URL on the staging domain; may not submit any form that sends data to an external service; may not click any link outside the app."
Record those limits as Decisions and rules in fenbs. A connected assistant reads them at the start of the session. If the task requires going beyond those limits, the assistant asks rather than proceeding. That keeps the human in the loop without requiring you to supervise every click.
The codex cli best practices guide covers approval policies in more detail, including which mode to use for different levels of risk.
Handle what needs review after the session
After a computer-use session, review the outcome the same way you would review a diff: did it do exactly what you asked, nothing more? Screenshots or a session log, if available, help confirm the steps the agent took.
Record the result on the task. If the agent completed the goal and you confirmed it, mark it done. If it completed part of the task or encountered an unexpected state, note what happened and what the next step is. That record is more useful than a session transcript because it says what was verified, not just what happened.
Keep approved limits where every session can read them
Record computer-use boundaries as Decisions and rules in fenbs: which domains the agent may visit, which actions require your approval, and what it may not do. Every connected session reads those rules before it starts, so you set the limits once instead of repeating them each time. Create a free account and add your project.
Create a free accountFree for one person · No credit card · Works with Claude, ChatGPT, Cursor and more
How to do this with fenbs
- Create a free account and open the project you are running computer-use tasks against.
- Write the action limits as a Decision in fenbs: what the agent may do, what it must ask about, and what it may not do at all.
- Connect Codex CLI to fenbs from Integrations. At the start of each session, ask it to read the project rules before it does anything.
- After each session, record the outcome on the task: what was confirmed, what was not, and any unexpected behavior to watch for next time.
- If a session reveals a gap in the limits, update the Decision before the next session runs.
For tasks that involve editing code as well as checking the result visually, the codex cli worktree guide explains how to keep the coding and the review steps in separate, isolated sessions.