OpenAI just turned Codex into a desktop agent
OpenAI just pushed Codex out of the “coding assistant” box and into full desktop-agent territory. It can now click, type, browse, generate images, remember preferences, and keep work going in the background, which mak...
Meta description: OpenAI’s Codex update adds desktop control, an in-app browser, image generation, and memory. Here’s why it matters.
What OpenAI announced
OpenAI says Codex is getting a major update that lets it operate your computer alongside you, work across everyday apps, generate images, remember preferences, learn from previous actions, and take on repeatable work. The desktop app also gets deeper support for developer workflows, including PR review, multiple files and terminals, SSH access to remote devboxes, and an in-app browser.
The practical shape of this is simple. Codex is no longer just a place to ask for code. It is becoming an agent that can move through the same tools humans use, and keep moving even when the task stretches over hours or days.
Why the desktop move matters
The big shift here is control. If a model can only suggest code, it is useful. If it can open apps, inspect pages, comment directly on a browser, and keep working in the background, it starts to look like an operating layer.
That matters because most real software work is not just writing code. It is bouncing between docs, tickets, terminals, browsers, screenshots, and weird little manual steps that never make it into the happy path. OpenAI is clearly trying to eat more of that surface area.
What changed inside Codex
OpenAI’s release bundles a lot together:
- background computer use on macOS
- an in-app browser with direct page comments
- image generation and iteration with gpt-image-1.5
- more than 90 plugins and integrations
- GitHub review comment handling
- multiple terminal tabs
- SSH to remote devboxes
- richer file previews for docs, slides, sheets, and PDFs
- a summary pane for plans, sources, and artifacts
- automations that can resume existing threads
- scheduled future work
- a preview of memory for preferences and prior context
- proactive suggestions for follow-up work
That is a serious bundle. It is also a clear sign that OpenAI thinks the next battleground is not just code completion, but task completion across the whole dev workflow.
| Capability | What it enables | Why it matters |
|---|---|---|
| Desktop control | Click, type, and use apps in the background | Codex can work where APIs do not exist |
| In-app browser | Comment directly on pages | Faster frontend, web, and game iteration |
| Memory | Recall preferences and prior corrections | Less repeated setup, less prompt debt |
| Plugins | Reach into tools like Jira, GitLab, Microsoft Suite | More of the workflow stays inside Codex |
| Automations | Resume tasks later and wake itself up | Better fit for long-running work |
How this compares with Anthropic and Google
OpenAI’s move lands right in the middle of the coding-agent arms race.
Anthropic has been pushing Claude Code hard, and the market has been treating it like the default choice for a lot of serious agentic coding work. OpenAI is now answering with more surface area, not just a better model.
Google is also in the mix with Gemini on Mac and broader workspace-style features, but Codex is the more pointed shot here. It is aimed straight at the “agent as coworker” category, not just general assistant use.
The important comparison is this, Anthropic has had momentum in coding, OpenAI is trying to turn that into a platform fight.
What Labs should watch next
-
Watch adoption, not just launch copy. If teams actually move real work into Codex, this stops being a feature story and becomes a workflow migration.
-
Watch the macOS constraint. Background computer use starts on Mac, which is fine for now, but it limits how fast this spreads.
-
Watch the memory layer. Persistent preference memory is where the product gets sticky, and where the risk profile changes too.
-
Watch the plugin ecosystem. The winner here may be whoever becomes the easiest place to connect tickets, docs, chats, and code.
-
Watch the response from Anthropic. This is a direct challenge. A reply is likely.
FAQ
Is this just another coding assistant update?
No. OpenAI is adding computer use, memory, browsing, and workflow integrations. That is a broader agent play.
Why is background computer use important?
Because it lets Codex work in apps that do not expose APIs, which covers a lot of real-world development tasks.
Does this beat Claude Code?
Not automatically. But it narrows the gap and makes the category more competitive.
What is the biggest product risk?
Trust. The more Codex can do, the more users will care about control, reliability, and what it remembers.
CTA
We are tracking the agent stack shifts that actually change how people build. Read more Labs coverage and keep an eye on the next Codex move.