Codex Desktop for Business: Setup, Approval Modes and Scheduled Tasks

Install and First Run

1. Download from openai.com/codex. On macOS, drag to Applications; Windows follows its normal installer flow.
2. Launch and sign in with a ChatGPT account, or create one. Account creation requires phone verification — he attributes this to spam mitigation.
3. Pick a workspace if your email belongs to more than one.
4. Onboarding offers to import your existing work from other agent apps, Claude Code and Claude Cowork included. If you're evaluating Codex alongside an existing setup, this is the step not to skip — see our platform comparison guide.
5. Create a project. A project is simply a folder on your computer, created for you when you name it; everything in that session is scoped to it.

The Settings That Actually Matter

Approval mode

Three positions, shown next to a shield icon:

• Approve individually — prompts every time the agent attempts an action it considers unsafe.
• Approve for me — the middle ground he recommends if you're undecided.
• Full access — unrestricted read/write across your machine. The UI labels this "warning: full access".

Model and effort

The picker exposes the models available on your plan — 5.6 Terra, Luna 5.5, and 5.4 mini in his account — alongside a separate effort control (medium, high, ultra, extra high). The tradeoff is stated plainly: higher effort consumes more of your token allotment, and running out mid-work is the failure mode that actually hurts. His default is medium, escalating only for genuinely hard tasks or when output quality disappoints.

Dictation

Built-in dictation processed locally through Whisper. He uses it for most of the course on the arithmetic that people speak at 150–200 words per minute versus typing at 50–75.

Features Worth Knowing About

• Scheduled tasks — the feature he calls central to using Codex for business. Tasks run autonomously on intervals you define. Shipped defaults include a weekday morning brief (calendar, unread email, priorities), a Friday weekly review of work done, and a follow-up monitor that flags email and calendar items needing attention. Custom examples he suggests: poll a mailbox every five minutes, watch a Slack channel and act on client messages as they arrive, or draft a staff member's work at the start of each weekday.
• Attach apps as context, not just files — beyond files and folders, you can attach a running application. His example is OBS Studio: import it and Codex can drive the recording software directly.
• Record a skill — Codex records your screen while you perform a task and builds a skill from it, rather than you writing the skill by hand.
• Plan mode — has Codex write dense planning material before committing to any change.
• Plugins and connectors — computer use (Codex controls your mouse and clicks for you), Chrome control, plus Notion, ClickUp, Dropbox, Asana and Linear. Connecting one hands Codex the keys to that account.
• Built-in browser and terminal — the right-hand panel spawns a Codex-managed browser (so it can view the web app it's building) and a terminal.
• Fork a conversation from any point in the thread.
• Parallelisation — he describes spawning ten or more concurrent conversations as core to getting value out of the tool, with the activity view surfacing only the ones needing attention.

Reading the Diff Panel

A section aimed at non-developers that's worth a line here because it's where new users get confused. Each turn shows counts like +90 and +116 per file. A first pass creating index.html and styles.css adds lines only. Asking for a change — his example is switching to a dark theme — produces +7 / -7 instead, because the agent locates the light-mode lines, removes them, and adds dark-mode ones. His framing: you won't read every change, but the panel tells you which files on your machine the agent touched, which is the accountability record you want when it gets something wrong.

Gotchas & Caveats

• Connectors are credential grants. Importing Notion, ClickUp, Dropbox, Asana or Linear gives the agent the ability to make requests and change things in that account on your behalf. Grant them one at a time and only where you need them.
• Computer use is a large escalation. An agent that controls your mouse can click anything visible on your screen, including confirmation dialogs. Treat that as a separate decision from letting it edit files.
• Effort costs allowance, not just time. Ultra and extra-high burn through usage limits meaningfully faster.
• Verify factual output about people and companies. His own demo — a landing page about himself — got his location wrong, his subscriber count wrong, and his age badly wrong. Web research from an agent needs checking.
• Promotional framing. The course is free and the instruction is genuine, but it opens with revenue claims and client name-drops. Read the business results as marketing and the interface walkthrough as the substance.

Key Takeaways

• Codex's onboarding can import an existing Claude Code or Claude Cowork setup, which makes side-by-side evaluation cheap.
• Approval mode is the most consequential setting in the app, and the three positions differ enormously in blast radius.
• Effort is a separate dial from model choice, and it spends your usage allowance rather than your wall-clock patience.
• Scheduled tasks are the feature that turns Codex from an assistant into unattended automation — and the one that most repays careful permission scoping.
• Skills can be recorded from screen activity instead of written, which lowers the barrier for non-developers considerably.
• Projects are just folders, so scoping an agent to a directory is the whole isolation model.