OpenAI just turned Codex into a desktop superapp
Today OpenAI shipped the biggest Codex overhaul since launch. Background computer use that clicks and types across your Mac alongside you. An in-app browser you can comment on to steer the agent. More than 90 new plugins. Native image generation. Persistent memory. And automations that wake up across days or weeks to keep long-horizon work moving.
OpenAI says more than 3 million developers already use Codex weekly, and the release pushes it well outside the coding niche it started in. TechCrunch framed the launch bluntly: "OpenAI takes aim at Anthropic with beefed up Codex that gives it more power over your desktop."
Here's the context that makes the timing matter. OpenAI shipped this the same morning Anthropic launched Claude Opus 4.7, its newest frontier model with big vision and coding gains. Anthropic spent the last few weeks making the model smarter. OpenAI spent them making the agent do more. The message: the AI coding wars moved out of the IDE. For more on that, read our coverage of it here.
Developer Dan McAteer called the new Codex a "superapp" that does what only humans could a year ago. After spending time with what OpenAI shipped, that framing holds up.
This is a huge release, so it's going to take us some time to dig through this in depth. Here are the highlights, but expect more coverage as we dive in to play with it.
First up, the TL;DR
Here's what happened
- Background computer use. Codex now sees, clicks, and types across your Mac with its own cursor. Multiple agents can work in parallel without interfering with what you're doing in other apps.
- In-app browser. You can comment directly on any webpage to instruct the agent, with auto-captured screenshots and DOM elements (the underlying page structure) for precise context.
- 90+ new plugins. Atlassian Rovo (for JIRA ticket management), CircleCI, CodeRabbit, GitLab Issues, Microsoft Suite, Neon by Databricks, Remotion, Render, and Superpowers, plus dozens more.
- Native image generation via gpt-image-1.5 for mockups, product concepts, and frontend designs inside the same workflow.
- Memory (preview). Codex remembers preferences, corrections, and context across sessions so you stop re-explaining the same things every time.
- Scheduled automations. Tasks can now wake up across days or weeks, re-using existing conversation threads so context carries forward.
- Proactive suggestions pulling context from Slack, Notion, Google Docs, and your codebase to propose what to work on first each morning.
- Software-lifecycle support: PR review comment handling, multiple terminal tabs, SSH to remote devboxes (in alpha), rich previews for PDFs / spreadsheets / slides / docs, and a new summary pane tracking agent plans, sources, and artifacts.
Why this matters: Codex used to live inside your editor. The new version lives on your desktop, on the web, inside your apps, and on its own schedule. Background computer use means agents can work in any Mac app, including ones without an API. The in-app browser turns the open web into an instruction surface. Persistent memory plus scheduled automations means Codex no longer needs you sitting at the keyboard to make progress. Box CEO Aaron Levie put it plainly: the new Codex lets knowledge workers offload drafting reports, setting up data rooms, reviewing contracts, onboarding clients, generating marketing assets, and processing invoices.
Our take: Codex and Opus 4.7 shipped on the same day, and the side-by-side tells the whole story. Anthropic focused on making the model smarter. OpenAI focused on making the agent do more inside the app. Both are correct moves. Both are incompatible with the assumption that AI coding tools stay politely inside your IDE. If you build software (or SaaS) for a living, today you got a much clearer picture of what the frontier labs are willing to compete for. Spoiler: the answer is everything.
What actually shipped
Background computer use
The headline feature. Codex can now see, click, and type across your Mac with its own cursor. Multiple agents can work on your machine in parallel without hijacking what you're doing elsewhere.
This matters because it removes the API dependency. Before today, Codex could only act inside apps that exposed developer hooks. Now it can use anything on your desktop: testing apps that don't have APIs, iterating on frontend changes, driving apps your company built five years ago and never wrote integrations for. OpenAI is pitching this as helpful for developers. The ceiling is obviously higher.
OpenAI's Alexander Embiricos described the implementation as "deep OS-level wizardry," because Codex can see, click, and type in apps in the background without taking over your machine (you can work in other windows in parallel). Embiricos also confirmed Windows support is coming, which closes the biggest question developers had about the launch.
OpenAI designer Ari Weinstein highlighted the work that went into the cursor itself, calling out natural, aesthetic motion paths built by his team (@supercgeek, @kierajmumick, @dexterleng, @philzet). Tiny detail, big UX lift. An agent cursor that jerks across your screen feels like malware. One that moves like a thoughtful person feels like a coworker. That choice was intentional.
Computer use is rolling out on macOS first, with EU and UK availability coming soon and Windows following.
The in-app browser (with comment mode)
This one might be the sneakiest upgrade in the release. Codex now ships with a built-in browser, and you can comment directly on any webpage to give the agent instructions. Codex PM James Sun demoed it: click, comment, and Codex auto-captures a screenshot plus the relevant DOM elements so it has precise context for the task.
For frontend work, this is a huge step up from "here's a screenshot, now guess which div I meant." For docs and research, it turns any webpage into a canvas the agent can read and act on with you. OpenAI says they plan to extend this "over time so Codex can fully command the browser beyond web applications on localhost," which is the kind of casual line that rearranges entire product roadmaps.
90+ new plugins
Codex's plugin system now bundles skills, app integrations, and MCP servers (Model Context Protocol, a standard for giving agents access to tools) into one surface. OpenAI highlights the ones developers will reach for most:
- Atlassian Rovo for JIRA ticket management
- CircleCI for CI/CD pipelines
- CodeRabbit for AI code reviews
- GitLab Issues
- Microsoft Suite (Word, Excel, PowerPoint, Outlook)
- Neon by Databricks
- Remotion for programmatic video
- Render for deployments
- Superpowers for custom workflows
Codex PM Derrick Choi broke down his 11 favorite new capabilities in a thread: projectless threads, artifacts for PDFs / docs / slides / spreadsheets, memories preview, remote SSH devboxes, in-app browser comment mode, single-threaded automations, and more. Worth reading end-to-end if you want a power-user's tour.
Image generation, directly in-flow
Codex can now call gpt-image-1.5 natively to generate and iterate on images. Combined with screenshots and code, that means mockups, product concepts, UI variations, and game assets all happen inside the same session instead of bouncing between five tools. For frontend and game developers specifically, this closes a gap that felt much bigger than it should have.
Memory and scheduled automations
Two features that quietly change the feel of using Codex.
Memory (in preview) lets Codex remember useful context from previous experience: your preferences, corrections you made last time, information it took hours to gather. "This helps future tasks complete faster and to a level of quality previously only possible through extensive custom instructions," OpenAI writes. Translation: fewer context prompts, less repetition, agents that feel like they actually know you.
Automations got a bigger rewrite. Codex can now schedule future work for itself and wake up automatically to continue on long-running tasks, potentially across days or weeks. Even better: automations can re-use existing conversation threads, preserving context from previous runs. OpenAI specifically calls out teams using these for landing open pull requests, following up on tasks, and staying on top of fast-moving conversations across Slack, Gmail, and Notion.
Codex also proactively proposes where to pick up each day, using context from your projects, plugins, and memory. The OpenAI example: Codex identifies open comments in Google Docs that need your attention, pulls relevant context from Slack, Notion, and your codebase, then gives you a prioritized list of what to do first. That's the first time a frontier lab has shipped what "start-your-day AI assistant" actually looks like in practice.
Support across the full software lifecycle
For the developers already using Codex inside their workflows, the update brings heavy lifting across the software development lifecycle (the whole process of writing, reviewing, testing, and shipping code):
- Addressing GitHub PR review comments
- Multiple terminal tabs in one workspace
- SSH to remote devboxes (in alpha), meaning Codex can work on your cloud development machines in addition to your laptop
- Rich previews for PDFs, spreadsheets, slides, and docs
- A new summary pane tracking agent plans, sources, and artifacts
Put together, these move Codex from "helpful autocomplete" to "shared workspace for writing, checking, reviewing, and shipping code."
The demos that sold the release
OpenAI's launch video tells you what the features are. The demos from the wider Codex team tell you what they feel like.
Nick Baumann, a Codex product lead, posted a fast-paced demo video showing how he uses the new Codex for "everything"; it rips through the full surface area of the release in a few minutes. Fellow team member Pashmerepat immediately flagged Baumann as the best follow for inspiring Codex demos, which tracks.
Dan McAteer shared his own demo spotlighting computer use specifically, and teased Spud (OpenAI's rumored next model) as the next shoe to drop. Worth noting for anyone trying to guess what else this year brings.
Aaron Levie, Box's CEO, argued in a thoughtful thread that the new Codex (computer use plus plugins) is a genuine step change in agent capabilities for knowledge workers. His list of what becomes automatable: drafting reports, setting up data rooms, reviewing contracts, onboarding clients, generating marketing assets, processing invoices. The new Box plugin, he notes, lets Codex automate almost any enterprise-content workflow. It's a useful reframe because Codex was positioned today as a coding tool; Levie is pointing out that the surface area is much larger.
One more: TestingCatalog reported earlier today that OpenAI was preparing Computer Use for Codex as an optional plugin with its own dedicated settings section. That confirmation came hours before the official launch. If you want to be early on what ships next, their scoops tend to land.
Even OpenAI's own staff couldn't stay quiet. Research engineer Tibo Sottiaux teased the release in the morning with a cryptic "Feeling codexy today." Later in the day, he celebrated Codex's 1-year anniversary by letting Codex reset its own rate limits across all plans (125K+ views and a small riot in the replies from users who'd just hit their weekly reset four hours earlier). A solid reminder that internal enthusiasm around a launch is almost always the best leading indicator; also, free tokens never go out of style.
Why this lands the way it does
Two bigger threads make the release hit harder.
Thread one: the AI coding wars are openly SaaS wars now. TechCrunch's framing ("OpenAI takes aim at Anthropic") sounds like a headline grab, but it's accurate. Anthropic CPO Mike Krieger resigned from Figma's board this morning as reports surfaced that Anthropic is shipping a design tool that would compete with Figma. TechCrunch called it "another data point for investors who fear the SaaSpocalypse, that the largest AI labs will come to dominate software businesses." OpenAI shipping computer use, plugins, memory, and automations all in one release is the same signal from a different angle. The labs aren't pretending anymore that their frontier models stop at the API layer.
Thread two: Sequoia's thesis just got a receipt. Sequoia partner Julien Bek argued in a piece published this week that the next $1T company will be a software company that sells the work, not the tool. His framing: writing code is intelligence, knowing what to build is judgment, and AI has crossed the intelligence line, so the big prize is selling outcomes (closing the books, drafting the NDA, brokering the insurance) at software margins. Codex with computer use plus plugins plus automations is literally that thesis in executable form. OpenAI isn't selling a coding tool. It's selling the work.
What to watch from here
A few things to keep an eye on:
- How many people install the plugins. OpenAI shipped 90+, but the real test is which ones get heavy use. Atlassian Rovo and Microsoft Suite are the early tell for enterprise traction.
- How quickly computer use expands. macOS-first today, EU and UK next, Windows confirmed as coming. Linux isn't mentioned, which is worth tracking if your dev stack runs there.
- Whether automations become trusted. Agents that wake up across days sound great until one makes a bad call at 3 a.m. The next few weeks of public anecdotes will tell us whether OpenAI got the guardrails right.
- What Anthropic does in response. Claude Code v2.1.111 already shipped today with its own upgrades, including a full desktop app redesign, parallel sessions across repos, SSH on Mac, and three view modes for orchestrating multiple agents. The overlap with what Codex just launched isn't a coincidence.
- Spud. McAteer's tease isn't nothing. OpenAI's pattern is to ship a product that's too good for today's model, then release tomorrow's model to actually run it.
The short version: today, OpenAI turned Codex into something closer to an operating-system-level agent than a coding tool. Anthropic did the same thing six weeks ago with Claude Code skills and Computer Use. The gap between "AI autocomplete" and "AI coworker" is closing fast, and the labs are done being subtle about it.
Go download the updated Codex app if you want to see it for yourself. Or wait a week and let the rest of the internet tell you what broke first. Either works.
Read more
Covered today in The Neuron's Around the Horn Digest alongside Claude Opus 4.7, Canva AI 2.0, Factory's $150M Series C, and everything else that shipped in AI before lunch.
