The only shortcoming lies in connection stability.

Recently, the pace of change with Codex has been remarkable. It started with an epic update for the desktop version, followed by remote SSH, then Computer Use and Chrome extensions. The number of weekly active users has now surpassed 4 million.

Today, OpenAI has integrated Codex into the ChatGPT mobile app. According to official statements, the mobile version of Codex is now available in Preview form on both iOS and Android versions of ChatGPT, supporting connections to macOS devices running Codex. Support for Windows is still pending.

However, the mobile version of Codex is more accurately described as an entry point to the desktop version of Codex on the phone. You can access Codex directly through ChatGPT on your phone, connect to the running environment on the desktop, and then initiate new requests on your phone for the desktop version of Codex to complete tasks.

Claude Code actually launched a similar Dispatch feature in March, allowing the Claude App on mobile to command Claude Code on the desktop. Earlier, OpenClaw (Longxia) had already achieved this by integrating with major IM platforms.

Better late than never. But more importantly, how well-developed is the mobile version of Codex? OpenAI, known for its product capabilities among the 'AI Big Three' in Silicon Valley, can it gain a late-mover advantage this time?

The Phone Becomes the Second Entry Point for Codex

It must be clarified that the mobile version of Codex is not a standalone new app.

In fact, OpenAI has not launched a separate Codex iOS or Android app this time. Instead, they have directly placed the entry point within the ChatGPT app. Opening ChatGPT on your phone allows you to see the 'Codex' entry in the sidebar.

Image Source: Leikeji

So, to some extent, you can understand the core of this update as allowing users to command Codex on their computers through ChatGPT on their phones.

Not only does this save over 1 billion ChatGPT users worldwide from installing an additional app, but more importantly, OpenAI has not made the mobile version of Codex a new development tool. Instead, they have incorporated it into the unified entry point of ChatGPT. For users, this means a short usage path. As long as the desktop and mobile versions are logged into the same account, ChatGPT on the phone can easily connect to the macOS desktop version with Codex installed.

I connected Codex on macOS using Codex on Android. Image Source: Leikeji

In actual use, the mobile version of Codex can already directly initiate new requests on the phone and then have the desktop version of Codex on the computer complete the tasks. At this point, the phone acts more like a lightweight task console. You clarify your requests, while the desktop version of Codex reads projects, executes commands, modifies files, and runs tests.

Meanwhile, the real-time status of the desktop version of Codex is also synchronously displayed on the phone, including thinking processes, operations, and results. Therefore, we can completely start commanding Codex to work on the computer and check progress on the phone when we go out, or directly guide execution and push long tasks forward without being tied down by devices.

Even models, speed, and reasoning levels can be directly adjusted on the phone.

Image Source: Leikeji

I think this division of labor design is very clever. Limited by factors such as environment, performance, and screen size, phones are not suitable for independently running complex tasks like Coding. They are more suitable as an 'anytime, anywhere' entry point for the desktop version.

The mobile version of Codex does not blindly transplant (Note: ' transplant ' is left as is for context, but it means 'port' or 'migrate' in English; however, since we're keeping HTML and not translating every word literally, it's acceptable here as part of the original text's tone), but instead picks out the most suitable tasks for mobile: initiating requests, viewing processes, approving operations, continuing conversations, and checking results. The complete thinking processes, operations, and execution results of the desktop version of Codex are all synchronously visible on the phone; the only thing not synchronously displayed is output files, which are only saved on the computer.

This handling is not surprising. After all, the files and running environment are originally on the local machine. The mobile version is only responsible for synchronizing status and interaction, so there is no need to duplicate all outputs to the phone.

In addition, the mobile version of Codex can connect to more than just one Mac. It can simultaneously connect to multiple desktop versions. For example, if both a MacBook Pro and a Mac mini are running Codex, one phone can manage the desktop versions of Codex on different computers. This also aligns with the real workflow of Agent products like Codex.

Image Source: Leikeji

Many developers do not just use one machine. They use a portable MacBook for daily projects and a Mac mini or remote environment for heavier tasks. These environments were often fragmented in the past, but now Codex is at least attempting to unify them into a single phone entry point. You can switch to different Hosts on your phone, view the progress of different tasks, and even distribute different requests to different machines.

However, there are still some obvious areas for improvement in the mobile version of Codex. For example, conversation editing is not yet supported. This issue may seem minor, but it is actually important in mobile scenarios. Phone input is more prone to typos, missing context, or realizing that a request was not fully stated just after sending it.

For an entry point that can remotely schedule tasks on the desktop, the inability to edit conversations makes people more cautious and reduces the enjoyment of initiating tasks on the mobile end. This indeed needs to be addressed in the future.

However, the biggest pain point I encountered this time was the frequent need to reconnect on the mobile end. Sometimes, just as you finish typing a request on your phone, the connection drops, and the message cannot be sent.

Even worse, sometimes the desktop version of Codex is still running, but the mobile end suddenly enters a reconnection state, and the interface briefly goes blank. Since the actual task execution occurs on the desktop end, this disconnection usually does not affect the running process and final results. However, from the perspective of mobile end interaction, this experience is indeed frustrating.

Image Source: Leikeji

Fortunately, such issues seem more like engineering stability problems in the Preview stage rather than product design problems.

In addition, the mobile version of Codex is not limited to serving the local desktop end. It also supports connecting to the Codex cloud, expanding the coverage of the phone entry point. Details like customizing pets in the desktop version can also be synchronized to the mobile end. The latter is certainly not a core capability, but it further illustrates that OpenAI is not treating the mobile version of Codex as a temporary remote viewing method but is genuinely building a new AI workflow and experience around Codex.

Viewing the mobile version of Codex within the entire Agent product landscape is even more interesting, especially regarding the understanding of Agent experiences in mobile scenarios, where there are still some differences.

On Mobile, Codex Differs from 'Longxia'

Nowadays, Agent products generally support mobile entry points, primarily to allow users to submit requests anytime, anywhere, and then have the Agent call tools, open web pages, process files, or complete processes, emphasizing 'turn to the Agent for anything.' This naturally applies to the mobile end, as a significant number of requests originate from mobile scenarios.

There are generally several approaches. One is open-source Agent tools like OpenClaw and Hermes, which basically support integration with IM, allowing users to submit requests to the Agent anytime, anywhere through commonly used chat apps. The popularity of various 'Longxia' earlier this year was largely due to this.

Image Source: Hermes Agent

Another approach is to extend the Agent workflow from the desktop to the mobile end, allowing users to continue monitoring tasks, issuing instructions, and making confirmations while outside. This is the approach taken by Claude Code and now Codex.

However, even when it comes to 'using Agents' on mobile, the experience focuses differ significantly. On OpenClaw or Hermes, you do not need to understand a complex new client or switch to a specialized development tool. You can simply submit requests to a bot in WeChat, Feishu, or Telegram, just like sending a message to a friend.

For ordinary tasks, this experience can even be smoother than opening a dedicated app. Especially for Agents that emphasize long-term memory, automation, and multi-tool invocation, IM entry points are naturally suitable for handling cross-time, cross-scenario tasks like 'remind me,' 'help me check,' 'follow up,' or 'execute a certain process.'

However, using IM as an entry point also has problems. Conversations can carry task requests but struggle to carry complex process states. Which tools the Agent is invoking, what step it has reached, which files have been modified, whether tests have passed, or whether a certain strategy is reasonable—all this information quickly becomes cluttered when presented through chat bubbles.

For general-purpose Agents, this may not be a fatal issue; but for AI Coding, the process itself is part of the product experience. You not only need to know 'it's done' but also how it was done, where it failed, and whether the next step should proceed.

In contrast, the mobilization of Claude Code and Codex is quite different. They did not choose IM Bots but instead started from the mobile apps of Claude/ChatGPT, which are more suitable for complex tasks than the chat entry points of Hermes or OpenClaw.

Image Source: Anthropic

However, the dual-end connection of Claude Code relies more on Dispatch, which is more like a feature built by combining the IM approach of OpenClaw or Hermes, allowing the Claude Code on the computer to be 'commanded' through the Dispatch feature page on the phone. In contrast, the mobile version of Codex directly synchronizes conversations from the desktop version, including real-time synchronization of thinking processes, operations, and results actually running on the desktop end.

In short, the switch between desktop and mobile scenarios with Codex is actually more seamless and natural, which may be why the occasional reconnection issues seem even more annoying currently.

In Conclusion

The mobile version of Codex is certainly not yet mature. Truly complex code reviews, architectural judgments, and high-risk changes still require developers to return to their computers. For ordinary users, this entry point is also unlikely to become as frequently used as ChatGPT chats in the short term.

But the direction is clear: OpenAI sees the phone more as a scheduling entry point for Agents. This judgment is crucial because while the capabilities of these AI Coding Agents are now beyond doubt, the next question is how to make these capabilities easier and more natural to access and use, making product interaction design the new key.

CodexOpenAIclaudeOpenClawHermès

Source: Leikeji

Images in this article come from: 123RF Royalty-Free Image Library