Codex in ChatGPT Mobile: coding from your phone while the agent works on your machine
🔎 OpenAI turns your phone into a command center for coding agents
On May 14, 2026, OpenAI announced the integration of Codex directly into the ChatGPT mobile app, on iOS and Android. The preview is rolling out for all plans, including Free and Go, in all supported regions. In practice, your Codex session runs on your Mac, your devbox, or any remote environment, and you control it from your couch, public transit, or a café terrace.
This is a paradigm shift. Until now, coding agents kept you cloistered in front of your terminal. You would start a task, wait, and review the diff. With this integration, the execution model and the control interface are decoupled. The agent works in the background on your machine. The human supervises remotely. And it's available even on the free plan, which means OpenAI isn't looking to monetize this feature right now — it's looking to lock down the ecosystem.
The competition is reacting. Anthropic offers the Claude Code Agent View, a multi-session dashboard that remains anchored in the terminal. xAI launched Grok Build, its coding agent CLI. But none have yet taken the leap into native mobile control via a mainstream app. OpenAI is gaining a significant tactical advantage.
The essentials
- Codex has been accessible from the ChatGPT app on iOS and Android in preview since May 14, 2026, for all plans (Free, Go, Plus, Pro, Team).
- The coding agent continues to run on your remote machine (Mac, devbox, server) via SSH, and you supervise it from your phone in real time.
- OpenAI claims HIPAA compliance for managed environments, directly targeting healthcare companies and regulated sectors.
- Codex's mobile UX is considered the best on the coding agent market in 2026 according to Apidog's comparison.
- This feature accelerates the adoption of coding agents by removing the constraint of being physically present in front of a terminal.
Recommended tools
| Tool | Main use | Price (May 2026, check on openai.com) | Ideal for |
|---|---|---|---|
| Codex dans ChatGPT Mobile | Remote control of coding agents | Free (preview, all plans) | Devs who want to supervise Codex outside the office |
| Claude Code Agent View | Multi-session dashboard in the terminal | Included with Claude Code | Devs rooted in a pure terminal workflow |
| Grok Build | xAI CLI coding agent | Free (preview) | Devs curious about the xAI ecosystem |
How Codex Mobile actually works
The architecture is simpler than it looks. Codex doesn't run on your phone — that would be absurd for an agent that needs to navigate your codebase, modify files, and run tests. The agent runs on your development machine, exactly as before. What changes is the control channel.
You have an active Codex session on your Mac. This session exposes an endpoint that the ChatGPT mobile app can reach via a secure tunnel. From the app, you can see the progress status, modified files, and console outputs. You can approve an action, reject it, or give a new directive. Everything happens in real time, according to the details shared by OpenAI in its official announcement.
Desktop environment synchronization is handled seamlessly. If you start a session on your Mac at the office, you can pick it up on your iPhone on the subway without losing any context. The New Stack confirms that remote SSH access is at the heart of this integration, with a particular focus on connection security.
It is this separation between the execution plane (the machine) and the control plane (the phone) that makes all the difference. You are no longer hostage to your terminal.
Availability and plans: even for free, the same battle
This is the most surprising detail. The Codex Mobile preview is not reserved for Plus or Pro subscribers. How2Shout and AlternativeTo confirm that all ChatGPT plans have access to it: Free, Go, Plus, Pro, and Team. In all regions where ChatGPT is available.
Why offer this for free? Two hypotheses. First, the preview serves as a large-scale testing ground. OpenAI needs real usage data to stabilize the feature before a general launch. Second, it's an ecosystem play. A developer who gets into the habit of supervising Codex from their phone develops a dependency on the OpenAI workflow. When the final version becomes paid or limited to certain plans, the switching cost will be high. It's soft lock-in, and it's well played.
The rollout began on May 14, 2026, according to 9to5Mac, with a progressive rollout on iOS and Android.
Codex Mobile vs Claude Code Agent View : two philosophies of control
Anthropic hasn't been sitting idle. The Claude Code Agent View also offers multi-session monitoring, but with a fundamentally different approach: everything stays in the terminal. You get a text-based dashboard capable of displaying multiple agent sessions in parallel, navigating between logs, and approving or rejecting actions. It's powerful, but you are stuck at your keyboard.
| Criterion | Codex Mobile (OpenAI) | Claude Code Agent View (Anthropic) |
|---|---|---|
| Interface | Native mobile app (iOS/Android) | Dashboard in the terminal |
| Remote control | Yes, from anywhere | No, requires an active terminal |
| Multi-session | Yes, via the app | Yes, split view in the terminal |
| Free plans | Yes (preview) | Via free Claude Code |
| Underlying models | GPT-5.3 Codex, GPT-5.5 | Claude Opus 4.7, Claude Sonnet 4.6 |
The comparison of the best LLMs for coding shows that GPT-5.3 Codex (agentic score: 80) and Claude Opus 4.7 (94.3) aren't playing in the same league when it comes to pure reasoning. But for agentic coding, the model is only part of the equation. The control interface and the ability to supervise without being physically present change the game in everyday use. A slightly less brilliant agent that you can pilot from your pocket is often better than a genius agent that keeps you glued to a screen.
If you are hesitating between the two ecosystems, the Claude vs ChatGPT comparison remains the best starting point for understanding the philosophical differences between Anthropic and OpenAI.
The changing workflow: launch, leave, supervise
The concrete value of Codex Mobile is measured by the scenarios it unlocks. Let's take a real-world example. You kick off a refactoring of a critical module at 6 PM. Before, you would wait in front of your screen, or kill the session to go home. Now, you launch Codex on your office machine, take the subway, and follow the progress on your iPhone.
You see that the agent has modified 14 files. You approve the batch. You notice it's hesitating on a unit test — you can inject a contextual hint from your phone to unblock it. You get home, the refactoring is complete, the tests pass. You spent only 5 active minutes on this task that would have required 45 with classic supervision.
This is exactly the principle I explored in my article about my AI agent that works while I sleep. The ability to delegate and supervise asynchronously is the real turning point for agentic AI in 2026. It's no longer automation, it's delegated management.
For teams, the implication is clear: the best autonomous AI agents are becoming full-fledged collaborators who no longer require your physical presence to move forward.
Security, SSH, and HIPAA compliance
Remote coding involves routing sensitive data — your codebase, your API keys, your secrets. OpenAI made sure to address this head-on. According to The New Stack, the Codex Mobile integration relies on end-to-end encrypted remote SSH connections. Your phone does not store the code. It simply relays control commands and displays states.
More notably, OpenAI claims HIPAA compliance for managed environments. This means that healthcare companies, subject to strict regulations regarding patient data, can theoretically use Codex Mobile to work on code that touches regulated systems. This is a strong signal to businesses: it's not a toy, it's a production tool.
One gray area remains. HIPAA compliance applies to environments managed by OpenAI for enterprise customers. If you use Codex Mobile on the free plan with your personal Mac and a homemade SSH tunnel, you are not covered. The distinction is important and often misunderstood.
The models behind Codex: GPT-5.3 Codex and GPT-5.5
Codex Mobile is not a model. It's an interface. The engine underneath is GPT-5.3 Codex, with an agentic score of 80 on the reference benchmark. This is sufficient for the vast majority of coding tasks — refactoring, test generation, debugging, implementation of specified features.
For more complex tasks requiring deep reasoning, you can configure your Codex session to switch to GPT-5.5, OpenAI's flagship model with an agentic score of 98.2. The extra cost is real but justified when the agent has to navigate a codebase of several hundreds of thousands of lines with cross-dependencies.
The choice of model remains an important lever. If you want to understand the nuances between model families, the ChatGPT vs Gemini comparison sheds light on the differences in approach between OpenAI and Google regarding chain-of-thought reasoning. And for the curious who want to test without spending money, the best free LLMs now include agentic options worth checking out.
Grok Build and the coding agent war
OpenAI is not alone in this space. xAI launched Grok Build, its first coding CLI agent, powered by Grok 4.1 (agentic score: 79). The approach is classic: a command-line agent that reads your codebase, suggests changes, and waits for your approval. No mobile interface, no remote supervision. It's functional, but it's the 2025 model.
The coding agent market is segmenting into two distinct layers. The execution layer (the model that codes) and the control layer (the human-agent interface). OpenAI is the first to understand that true differentiation wouldn't happen on the model — scores are all compressing upwards — but on the control interface. Putting this control layer in the pockets of 200 million ChatGPT users is a major strategic move.
For those who want to go further and build their own agentic workflows, the guide on how to create an AI agent details the possible architectures. And if you prefer to keep total control locally, open source AI agents with Ollama offer an alternative without cloud dependency.
Agentic memory: the missing link
One point that OpenAI's announcement glosses over: context persistence between mobile sessions. You monitor Codex from your phone in the evening. The next morning, you reopen the app. Does the agent remember what happened? Your approvals, your rejections, the task context?
The question of AI memory is central to the adoption of long-term agents. A coding agent that forgets why it made a certain architectural choice in the previous session reverts to being a generic assistant, not a persistent collaborator. The best LLMs for AI agents are those that handle long-term memory well — and that isn't necessarily the one with the best benchmark score.
For now, Codex Mobile inherits the context management of the desktop session. If your Codex session on Mac is still active, the context is intact. If it has been closed, it's a new session. OpenAI will need to address cross-session persistence for the "mobile-first" model to deliver on all its promises.
Real cost: how much it costs to run an agent 24/7
The accessibility of the preview hides a question that every dev will ask: how much does a Codex agent running continuously cost? The answer depends on the plan and the model chosen.
On the Free plan, you have access to Codex with GPT-5.3 Codex, but with usage limits that have not been publicly detailed by OpenAI. On the Plus ($20/month) and Pro ($200/month) plans, the limits are more generous. The real cost is not in the ChatGPT subscription — it is in the computing power of the machine running the agent.
If you use an M4 Pro Mac on standby with Codex, the power consumption is negligible. If you rent a cloud devbox, add €30 to €100/month depending on the configuration. The article on my AI agent that works 24/7 for €29/month details this arithmetic precisely. The total remains well below the cost of a junior developer for the same repetitive tasks.
To host the front-end of your projects managed by Codex, Hostinger remains an economical option at under €3/month for basic needs.
Impact on the adoption of coding agents
Coding agents have existed since late 2024. Cursor, Claude Code, Codex, Grok Build — the market is saturated with capable tools. But the actual adoption rate remains modest. Why? Because the user experience required sitting in front of a screen, in a terminal, with continuous attention. It was automation with a human in the loop, but a tight loop.
Codex Mobile loosens this loop. By making control asynchronous and mobile, OpenAI eliminates the main friction to adoption. You no longer need to block out an hour in front of your screen. You launch it, you go about your life, you step in when necessary. It's the shift from a "copilot" mode to a "supervisor" mode. And this semantic shift reflects a real change in practice.
Adoption data will be interesting to follow. If OpenAI releases usage figures in the coming months, I bet that Codex sessions controlled via mobile will represent more than 40% of the total by the end of 2026. The phone is the most accessible interface in the world. Making a coding agent accessible from this interface means making it accessible to everyone.
❌ Common mistakes
Mistake 1: Thinking Codex runs on the phone
The mobile app is a thin client. It sends commands and displays statuses. All the heavy lifting — reading the codebase, executing, file modifications — is done on your remote machine. If your Mac is turned off, Codex does nothing, regardless of what you do on your phone.
Mistake 2: Ignoring SSH tunnel security
By default, Codex Mobile sets up an encrypted tunnel. But if you are on a public Wi-Fi network and you haven't enabled strong authentication on your session, you are exposing an attack vector. An attacker who intercepts your mobile session could theoretically approve malicious modifications on your machine. Always enable two-step verification on your OpenAI account and verify that the SSH tunnel uses keys, not passwords.
Mistake 3: Approving without reading the diff
The ease of mobile approval is a trap. Swiping a finger to approve 14 modified files is tempting when you're on the subway. But every approval is a commit in your codebase. Take 30 seconds to read the modification summary before validating. Codex makes mistakes, especially with GPT-5.3 Codex which is not OpenAI's most robust model.
Mistake 4: Confusing the free preview with production-ready
This is a preview. Display bugs, untimely disconnections, desynchronized states between mobile and desktop — this is expected. Do not launch a critical database migration with mobile supervision only. Keep a desktop access as a fallback.
❓ Frequently Asked Questions
Does Codex Mobile replace Claude Code ?
No. These are tools with different philosophies. Claude Code excels in terminal-heavy workflows with deeper reasoning via Claude Opus 4.7. Codex Mobile excels in remote supervision. Many devs will use both depending on the context.
Is it really free ?
The preview is free for all ChatGPT plans. OpenAI has not communicated about the pricing of the final version. It is likely that some advanced features (multi-session, premium models) will be reserved for paid plans.
Can Codex Mobile be used with a Windows PC ?
OpenAI's announcement and the articles from MacRumors focus on Mac. The SSH protocol theoretically works with any machine, but the native integration is currently optimized for macOS. Windows support should arrive, but no timeline has been announced.
What is the latency between an action on the phone and its execution ?
Early user feedback suggests a latency of 1 to 3 seconds for displaying states, which is perfectly acceptable for supervision. The execution itself depends on your local machine and the complexity of the task, not the mobile channel.
Can multiple Codex sessions be launched from the mobile ?
The current mobile interface seems limited to one active session at a time, unlike Claude Code Agent View which displays multiple sessions in parallel in the terminal. This is a limitation that should be lifted in a future update.
✅ Conclusion
Codex in ChatGPT Mobile isn't just an additional feature — it's OpenAI redefining the contract between the developer and their coding agent: execution remains local, control becomes mobile and asynchronous. Starting in May 2026, the question is no longer "which agent codes best", but "which agent can I supervise without being stuck in front of my screen". If you want to understand how to build a complete agentic workflow, the guide on the meilleurs agents IA autonomes is your next read.