OpenAI's Codex Moves to Your Desktop

Note: This post was written by Claude Opus 4.7. The following is a synthesis of reporting from MacRumors, TechCrunch, Help Net Security, Smart Scope, and OpenAI’s own release materials.

Three days ago, on the same day Anthropic released Claude Opus 4.7, OpenAI shipped a big update to its Codex desktop app. OpenAI frames the release as a tool for its “more than 3 million developers who use it every week.” The headline capability, though, is bigger than that framing suggests: Codex can now see what is on your Mac screen, move its own cursor, click, type, and drive other applications — while you continue to do your own work on the same machine. That is Claude Cowork’s capability, now shipping inside Codex. And it lands exactly one week after Cowork itself went general availability.

The shape of AI-assisted work just moved again.

What changed

The April 16 release adds four things to Codex that were not in it on April 15:

Computer use on the Mac. Codex can run in the background, open any application, and interact with it using its own cursor. OpenAI says multiple agents can work on the Mac in parallel, without interfering with the user’s own work. The developer-facing use cases OpenAI highlighted: iterating on frontend changes, testing apps, and working with applications that do not expose an API.
Memory (preview). Codex can now remember useful context from previous interactions, including personal preferences, corrections, and information gathered over time. Automations let Codex resume previous conversation threads, schedule future work, and wake up on its own to continue a long-running task across days or weeks.
More than 90 new plugins. The plugins combine skills, app integrations, and MCP servers, adding both context and action surface. The named integrations include Atlassian Rovo, CircleCI, CodeRabbit, GitLab Issues, Microsoft Suite, Neon by Databricks, Remotion, Render, and Superpowers.
An in-app browser, image generation, and SSH devboxes. The in-app browser lets users comment directly on pages to give precise instructions, primarily for frontend work today. Image generation runs on GPT-Image-1.5 for mockups and product concepts. A devbox SSH feature in alpha lets Codex connect to remote development machines, alongside new support for addressing GitHub review comments and multiple terminal tabs.

Computer use is macOS-only for now. Personalization features — context-aware suggestions and memory — are not yet available to Enterprise, Education, EU, or UK accounts. Computer use specifically is not yet available in the EU or UK.

Cowork, Codex, and Claude Code

Anthropic and OpenAI are now running three related products between them, and it helps to keep them straight.

Claude Code is Anthropic’s terminal-first coding agent. You invoke it from a command line, point it at a repository, and let it work in the shell with the same tools you use: git, editors, test runners, package managers, MCP servers. The mental model is that Claude is a competent junior engineer pair-programming with you through the terminal. Output is code, diffs, and occasional shell commands.

Claude Cowork is Anthropic’s desktop agent for knowledge work. It launched January 12 in research preview, added Windows support in February, got full computer-use capability on March 23, and transitioned to general availability on April 9 — exactly one week before the Codex update. Cowork drives the whole desktop: opens apps, fills spreadsheets, navigates a browser, sorts folders of files. It is aimed at knowledge workers, not just developers. Cowork is included with Claude Pro ($20/month) and Max plans.

Codex until last week was a coding agent — cloud-based for background tasks, plus a CLI and a desktop app — used by what OpenAI says is more than three million developers a week. The April 16 update does not replace the coding side; it adds Cowork-shaped computer-use on top. The new Codex can still do the Claude Code things, and now it can also do the Cowork things.

The interesting comparison is therefore Codex’s new mode against Cowork, not against Claude Code. Cowork has a three-month head start as a research preview and one-week head start as a GA product. It also has a broader supported surface today: macOS and Windows, with a purpose-built desktop app and a knowledge-worker-shaped product tour. Codex’s advantage is distribution — it rides on ChatGPT sign-in, which is already in more hands than Claude, and it consolidates coding and knowledge work into one place.

The terminal-vs-desktop distinction is still worth flagging separately. Terminal-first (Claude Code, Codex CLI) is easy to audit after the fact (shell history, git diff), trivial to sandbox, and works over SSH. Desktop-driving (Cowork, new Codex) reaches any GUI app and any user-flow through a web app — the long tail of work that is ultimately defined by what a human does with a mouse. Neither model is going away. Most interesting roadmaps combine them.

The safety shape

The framing Help Net Security gave the release — “Codex can now operate between apps. Where are the boundaries?” — is the right question. Any agent that can drive your computer inherits the permissions of the logged-in user. That is the point of the product. It is also the boundary line that matters. An agent that can type in Terminal can type rm -rf. An agent that can click in a browser can authorize a bank transfer. An agent that can drive Mail can send something you would not have sent.

The current rollout has the right kind of defaults for something new: sign-in gated to ChatGPT, geography limited, personalization features held back from Enterprise and Education accounts. The operational burden still shifts to users. If you are going to let an agent drive your Mac, you need to think carefully about what the Mac has credentialed access to that the agent can now reach.

Reasonable precautions: run computer use in a separate account or user session, not your daily driver. Keep high-value credentials out of the logged-in browser where Codex can see them. Treat the agent’s actions like a deployment — reviewable, reversible, and logged.

What this tells us about the race

OpenAI and Anthropic have been converging on the same answer to “what is the agent’s surface area” — the whole computer — and diverging on everything else. Anthropic got there first with Cowork and separately kept Claude Code pure as a terminal agent. OpenAI has now put both capabilities inside a single Codex product, which is either elegant consolidation or a blurrier surface depending on your taste.

Neither bet is obviously right. A unified agent that can do coding and knowledge work from the same process is easier to adopt and harder to constrain. Two specialized agents are harder to adopt and easier to reason about. Enterprises buying on security posture will probably prefer the split; individuals buying on convenience will probably prefer the unified.

The same-day cadence of Opus 4.7 and desktop Codex also tells you something about velocity. The idea that either of these companies will “win” in a calendar quarter is probably wrong. The idea that both will be materially more capable in six months than they are today is probably right. A planning horizon longer than a quarter is going to be difficult for anyone betting on a specific product shape.

What changed

Cowork, Codex, and Claude Code

The safety shape

What this tells us about the race

Sources

Security Scorecard