OpenAI has rolled out a major upgrade to its Codex system that turns the coding assistant into a far more powerful “agent” on your desktop, in a move widely seen as a direct response to Anthropic’s push with Claude Code and remote-computer control tools. The revamped Codex can now quietly operate in the background on a user’s Mac, open and control apps with a visible cursor, and coordinate multiple AI agents in parallel, deepening the race over who will own the next generation of developer desktops.
Codex started life as an AI coding assistant embedded in chat interfaces, where developers could ask it to write functions, fix bugs or propose pull requests from the cloud. OpenAI’s own documentation describes it as a system that “can perform tasks for you such as writing features, answering questions about your codebase, fixing bugs, and proposing pull requests for review”.
With this week’s release, Codex is no longer just a sidebar in a browser but a full desktop application and agentic layer designed to sit at the heart of a developer’s machine. OpenAI says the Codex app for macOS is “a command center for agents” that provides “a focused space for multi-tasking with agents”, organizing their work into separate threads and projects so users can switch tasks “without losing context”.
The Codex desktop app effectively turns the assistant into a central workspace: users can connect local repositories, assign tasks to agents, review diffs, and manage multiple coding threads at once. A recent explainer notes that “instead of using a browser interface, Codex now works as a dedicated development environment where you can manage multiple threads, automate coding tasks, review AI actions, and even run projects in local or cloud environments”.
The headline change in the latest update is Codex’s ability to actively control a user’s computer, rather than just modifying files in a sandboxed environment. “Codex can now operate in the background on your computer, opening any app on your desktop and carrying out operations with a cursor that clicks and types,” the company said in an update about the new capabilities.
Functionally, this means Codex can deploy multiple agents that each work on different tasks on a user’s Mac “in parallel, without interfering with your own work in other apps,” according to OpenAI’s description. In practice, the system can launch development tools, navigate settings menus, or run utility apps while the person at the keyboard continues doing something else.
OpenAI stresses that the app includes sandbox controls that limit which directories and network resources Codex can access, giving teams some guardrails over what the agent is allowed to touch. The feature list for the Codex app highlights that “the sandbox controls which directories and network access Codex can use,” a key point as questions about data security and AI autonomy grow louder in enterprise environments.
Alongside desktop control, OpenAI is also expanding Codex’s reach into the web and third‑party tools. The latest release adds an in‑app browser that lets users issue higher-level commands, which Codex then carries out on specific web applications. OpenAI says this will be especially useful for front‑end and game development and plans to “eventually expand the capability so that Codex can ‘fully command the browser beyond web applications on localhost’”.
Codex now supports more than 100 integrations and plugins, including issue trackers and code review tools, allowing agents to move more fluidly between coding, testing and project management within a single workflow. One summary of the release notes that the tool “features 111 plugin integrations (e.g., CodeRabbit, Gitlab Issues), an in-app browser for development, and new pay-as-you-go enterprise pricing”.
By folding these hooks into the desktop app, OpenAI is positioning Codex not only as a code-writing engine but as a broader automation layer that can submit issues, follow up on bugs, and manage engineering tasks end‑to‑end. Analysts say that kind of deep integration is aimed squarely at corporate customers who want fewer windows to manage and tighter control over how AI agents interact with internal systems.
The timing and nature of the update make it difficult to ignore the competitive backdrop. Only last month, Anthropic announced enhancements to its own coding product, Claude Code, and related Cowork tools that allow their AI to remotely control a Mac and desktop on a user’s behalf while they are away from the keyboard. As one report put it, “watchers of the AI coding space will also note that some of the powers OpenAI is now adding to Codex seem to resemble those previously released by Anthropic for Claude Code”.
Anthropic currently leads many coding benchmarks at the high end, with the Claude Opus 4.6 model and a 1 million‑token context window that appeals to teams working across large repositories and complex specs. A recent developer‑focused comparison noted that “Anthropic has the stronger top-end coding story today, especially with Claude Opus 4.6 and Claude Sonnet 4.6,” while OpenAI is “usually easier to justify on price for managed API use”.
For OpenAI, beefing up Codex with multi‑agent capabilities, desktop control and a richer plugin ecosystem is a way to close the feature gap in areas where Anthropic has staked out early ground. One analysis of the competitive landscape summed it up bluntly: OpenAI’s specialist GPT‑5.3 Codex model and Anthropic’s Opus 4.6 “now compete directly as coding specialists,” with the race increasingly defined by productivity features and tooling rather than raw language modeling alone.
The new Codex update also expands how and where people can interact with their desktop through AI. Community guides show developers turning Codex into a kind of remote AI operator, controlling the Codex Desktop environment from a phone: “From a mobile device, users could interact with their Codex Desktop environment sending instructions, reviewing outputs, and supervising actions,” one walkthrough notes.
That sort of remote operation is part of a broader shift toward what many in the industry call “agentic AI”, where systems don’t just answer questions but take sequences of actions on behalf of the user. OpenAI is leaning into that framing, describing Codex as a way to “run work in parallel” across multiple threads and keep “a live audit trail of diffs and comments before pushing code”.
The Codex app operates as a hub on macOS where users “create projects, spin up threads, run work in parallel and keep a live audit trail of diffs and comments before pushing code,” according to a recent enterprise‑focused overview. That same report highlights that Codex is available across paid and free tiers of OpenAI’s product line, “enabling many teams to experiment without additional procurement hurdles”.
Supporters argue that these upgrades could dramatically increase developer productivity by offloading routine tasks and allowing engineers to coordinate multiple AI workers at once. As one explainer put it, “users don’t have to write every line of code by hand; they can just say what they want, and Codex will create the right code,” with the desktop app designed to make it “easier to keep track of projects and workflows”.
Enterprise commentators, however, are already flagging trade‑offs. A widely shared summary of the update noted that while Codex’s new powers “could redefine task automation for professionals, promising efficiency,” they also “raise questions about system control and data privacy”. Those concerns are likely to intensify as Codex gains deeper file-system and browser capabilities, even with sandboxing in place.
The move also lands as regulators and large employers are still working out policies around AI agents that can click, type and browse on their own. Security teams will want to understand how commands are logged, how access is limited, and what happens if an agent is steered toward sensitive internal resources either by accident or through prompt manipulation.
For now, the upgraded Codex marks one of the clearest signs yet that AI coding tools are evolving into full desktop companions rather than isolated assistants. OpenAI’s own description of the revamp underscores that point, saying the “agentic update and other new additions demonstrate [its] desire to not only make Codex a competitive coding assistant but also a more multifaceted tool that can be integrated into a variety of corporate workflows”.
Anthropic, meanwhile, is betting on larger context windows and strong safety positioning to differentiate Claude Code and its associated tools. A detailed comparison from earlier this year framed the choice as a trade‑off: “OpenAI is smoother for broad tool ecosystems; Anthropic is stronger if long-context coding quality is the center of the purchase.”
Developers and IT leaders will now have to decide whether the convenience of letting Codex click and type around their desktops outweighs the unease that comes with handing that level of control to an automated agent. What is clear is that the battle over AI‑powered coding has decisively shifted from the browser tab to the operating system itself, and both OpenAI and Anthropic appear determined to make their agents feel at home there.
Comments