Codex Mobile Is Here. Claude Had It First. OpenClaw Had It Before Either of Them.

By Addy · May 15, 2026

On May 14 in OpenAI's U.S. announcement, which is May 15 for this publication, OpenAI launched Codex mobile: the ability to steer a Codex coding session from your phone, available to every ChatGPT user including Free and Go tiers. The announcement was clean, the feature is real, and the coverage has been positive.

There is a longer story underneath it.

Ten days before Anthropic shipped the same capability in February, OpenAI had quietly hired the Austrian developer who had already built it without enterprise guardrails, for a reported 3.2 million users, from a cafe in Madrid. His name was Peter Steinberger. His project was called OpenClaw. Sam Altman said he was joining to "drive the next generation of personal agents."

That was February 15, 2026. Ten days later, Anthropic launched Claude Code Remote Control. Today, three months after that, OpenAI launched Codex mobile.

Both features do something OpenClaw was already doing. Both companies know this. Neither has said so publicly.

What OpenClaw Actually Was

Before comparing the two mobile launches, it is worth understanding what they are both responding to.

OpenClaw was not a coding tool. It was a local AI agent framework that could operate across your entire computer: browser, files, applications, system interfaces, with minimal restriction and maximum surface area. The developer community loved it precisely because it had so few guardrails. You gave it a task, it figured out how to do it, and it did not stop to ask permission for every step.

OpenClaw reportedly reached 346,000 GitHub stars and 3.2 million users. The ClawHub skills marketplace had over 10,700 available community-built extensions. Developers were building entire workflows on top of it: CRM systems, research pipelines, cross-application automations, that no official AI product had made possible because no official AI product had been willing to give an agent that much access to a local machine.

The tradeoff was real. Security researchers audited ClawHub and found 824 confirmed malicious entries out of 10,700 available skills, with 335 traced to a single coordinated attack operation. More than 30,000 OpenClaw instances were found exposed on the public internet without authentication. The Moltbook social layer suffered a breach that exposed 1.5 million API tokens and thousands of private conversations.

The power came from the lack of guardrails. The danger came from the lack of guardrails. These were the same fact described from different angles.

OpenAI's own previous attempts at agentic products, including its Agents API, Agents SDK, and the Atlas agentic browser, failed to gain the traction that OpenClaw achieved seemingly overnight. So OpenAI did what large companies do when the open-source community builds something better than their internal teams: they hired the creator, sponsored the foundation, and absorbed the insight into their product roadmap.

The deal closed quietly, no press release, no fanfare. Just two posts. Steinberger joined. OpenClaw became a foundation. OpenAI got the engineer who had proven that people wanted AI that does things, not just AI that says things.

Claude Code Remote Control: The February Launch

Ten days after Steinberger joined OpenAI, Anthropic shipped Claude Code Remote Control.

Launched on February 24 in the U.S., February 25 for much of the world, Remote Control is a synchronization layer that bridges a local Claude Code CLI session with the Claude mobile app and web interface. The session keeps running on the developer's machine. The phone is a remote window into that session, showing conversation history, diffs, tool calls, and approval prompts in real time.

The architecture is deliberately local-first. Your code never leaves your machine. Only chat messages and tool results flow through an encrypted bridge. Files, MCP servers, environment variables, and project settings all stay local. No inbound ports open on the machine. All traffic flows through the Anthropic API over TLS.

From the phone, a developer can see exactly what Claude is doing in real time, approve or reject file changes, provide additional instructions, redirect the work if needed, and monitor sessions from another device.

The limitations at launch were specific and practical. Remote Control required a paid Claude subscription, ran through a local process that had to stay open, and could time out if the host machine lost network access for an extended period. Anthropic's current docs still describe the core tradeoff clearly: the web and mobile interfaces are a window into a session running on your own machine, not a cloud copy of your project.

The timing, ten days after the OpenClaw hire announcement, was not coincidental. Anthropic was watching the same developer appetite that had given OpenClaw millions of users and responding with a safer, more controlled version of the same idea. Stay in your terminal. Approve sensitive actions. Code stays local. The guardrails OpenClaw explicitly lacked are the features Anthropic explicitly built.

Codex Mobile: The May Launch

Today, OpenAI shipped the same concept on a different surface.

Starting with this rollout, all ChatGPT users including those on Free and Go tiers can control Codex from the ChatGPT mobile app on Android and iOS. The phone acts as an intermediary. You are not coding on the phone. You are steering a Codex session running on a desktop or remote environment. Files, credentials, and permissions stay on the machine where Codex is running.

OpenAI uses a secure relay layer that keeps trusted machines reachable across devices without exposing them directly to the public internet. That relay also keeps active session state and context synced anywhere you are signed in with ChatGPT.

The architecture has the same shape as Claude Code Remote Control. Local execution. Relay-based sync. Phone as interface, not as execution environment. The philosophical approach, code stays where the work is running and remote access flows through an authenticated bridge, is the same design decision made by two different teams, three months apart.

The differences that matter are not architectural. They are distributional.

Claude Code Remote Control launched for paid Claude subscribers. Codex mobile launched for every ChatGPT user including the Free tier. OpenAI chose broad access on day one. Anthropic chose subscription gating on day one. Both decisions reflect each company's current strategic priorities: Anthropic protecting the economics of its highest-value developer workflows, OpenAI using mobile access as a distribution play to close the gap in developer adoption.

The other difference is polish. Codex mobile launched with PR status badges built in: open, merged, closed, draft states visible from the mobile interface. It supports working across threads and approvals from the mobile app, with screenshots, terminal output, diffs, test results, and approvals flowing back in real time. Claude Code Remote Control has improved since launch, but its first version made the tradeoffs more visible.

OpenAI had roughly eleven weeks to watch what Anthropic shipped and decide where to improve it. The Codex mobile launch reflects that advantage.

The OpenClaw Comparison Nobody Is Making

Here is the comparison that the launch coverage is not making but should be.

Claude Code Remote Control and Codex mobile are both good products solving a real problem. They are also both significantly more restricted than the thing they are responding to.

OpenClaw gave you access to your entire machine: browser, filesystem, system interfaces, cross-application workflows. The phone interface was one of the things you could use to initiate and steer those workflows, but the agent's surface area was not limited to a coding terminal. It could do anything a human operator could do on that machine, constrained only by the instructions you gave it.

Claude Code Remote Control gives you access to a Claude Code coding session. The agent operates within the terminal context. The mobile interface shows you diffs and tool calls. You approve file changes and redirect the work.

Codex mobile gives you access to a Codex coding session running in whatever environment you have configured: local Mac mini, remote cloud environment, corporate infrastructure. The mobile interface lets you check progress, prompt the agent, and receive updates including screenshots and test results.

Both are narrower than OpenClaw. Both are safer than OpenClaw. Both have fewer guardrail failures than OpenClaw. Both have fewer capabilities than OpenClaw.

OpenClaw's power came precisely from the lack of guardrails that would be unacceptable in a corporate environment. The race to build the safe enterprise version of OpenClaw is now the central question facing every platform vendor in the space.

Claude Code Remote Control and Codex mobile are both answers to that question. Neither of them is OpenClaw. They are the sanitized, guardrailed, enterprise-acceptable versions of the thing OpenClaw proved people wanted. The developer who wanted full machine access with minimal friction will still reach for OpenClaw or its open-source successors over either of these. The developer who needs to operate within a corporate security policy, a regulated industry, or an enterprise compliance framework will reach for the version with guardrails.

This is not a criticism of either product. It is an accurate description of the tradeoff both companies made explicitly. The question is whether the enterprise-acceptable version of the insight captures enough of the value to make the restrictions worth accepting.

The Hire That Explains the Timeline

The full picture of this week's launch requires going back to February 15.

When OpenAI hired Steinberger and absorbed OpenClaw's insight, they did not just get an engineer. They got three months of market validation for a specific hypothesis: developers want AI agents that operate on their local machines with minimal friction, and they will accept significant security risk to get that capability.

The ChatGPT subscription integration that followed confirmed the commercial validation. OpenAI opened ChatGPT subscriptions to OpenClaw's reported 3.2 million users, allowing subscribers to run autonomous agents via GPT-5.4 for $23 per month. Anthropic made the opposite decision, blocking flat-rate Claude subscriptions from covering OpenClaw usage in April and pushing those workloads toward usage-based API access.

Two companies, same product, opposite bets. OpenAI chose distribution: onboard the existing OpenClaw user base by making their preferred tool run on GPT-5.4 for a price lower than Anthropic's highest developer plans. Anthropic chose protection: prevent OpenClaw's guardrail-light architecture and heavy autonomous usage from blowing through the economics of flat-rate subscriptions.

The security concern was real. OpenClaw had documented incidents. Anthropic's decision was not arbitrary. But the consequence was that developers who had built workflows on OpenClaw with Claude as the underlying model were told in April that the combination no longer worked under the same subscription terms. OpenAI offered them the same class of workflow with GPT-5.4 for $23 a month.

That context makes today's Codex mobile launch make more sense. OpenAI is not just closing a gap with Claude Code Remote Control. It is building the mobile infrastructure for an ecosystem that now includes OpenClaw's former user base, OpenClaw's creator, and OpenClaw's institutional knowledge about what developers actually want from local agent access.

The Innovation That Was Not

There is a framing problem with how both of today's mobile launches are being covered.

Codex mobile is being described as a new capability. Claude Code Remote Control was described as a new capability in February. Both are being positioned as innovations in developer tooling.

Neither of them invented the concept. Before Remote Control, developers cobbled together workarounds: SSH tunnels, tmux sessions attached from mobile terminals, ngrok proxies, and custom WebSocket bridges. These were functional but fragile, prone to timeout issues, and required significant technical setup.

Before those workarounds, OpenClaw had already demonstrated that a sufficiently motivated developer community would build exactly this capability themselves if the official products did not provide it. The 3.2 million users were not waiting for Anthropic or OpenAI to invent remote mobile agent control. They had already found it.

What Claude Code Remote Control and Codex mobile both provide is the native, secure, integrated version of a workflow that the open-source community had already validated. They are not innovations. They are productizations. The distinction matters because productization is its own kind of value: removing the fragility, the security exposure, the configuration overhead. But calling it innovation misattributes where the insight came from.

The insight came from Steinberger in a Madrid cafe in November 2025. Everything since then has been the enterprise-acceptable version of what he built.

Who This Is Actually For

The honest answer to "who should use Codex mobile versus Claude Code Remote Control" is that the choice depends less on the feature comparison and more on the rest of your stack.

If you are already on Claude Code for production work, enterprise, regulated industry, local-first architecture, Remote Control is the natural extension. Your MCP servers, your project configuration, your CLAUDE.md files all stay local and accessible. The mobile interface is a window into the session you already have. The guardrails are the same guardrails you have already accepted. The limitation is the limitation you have already paid for.

If you are on Codex, or if you are a developer on ChatGPT's Free or Go tier who has never been able to justify a paid Claude subscription, Codex mobile is accessible in a way that Claude Code Remote Control was not at launch. The Free tier inclusion is the detail that will drive the most adoption. A developer who would not pay for remote agent access might use it for free if it is already in the app they open every day.

If you are a developer who wants the full OpenClaw surface area, full machine access, cross-application workflows, maximum agent autonomy, neither of these products is the answer. The safe enterprise version of OpenClaw is narrower than OpenClaw by design. The products that replaced it for guardrail-light development are the open-source successors that emerged after Steinberger joined OpenAI and OpenClaw became a foundation.

The market has segmented. Enterprise developers get Claude Code Remote Control and Codex mobile, both with guardrails, both local-first, both architecturally sound. Power users and developers in environments where security policy is flexible get the open-source successors to OpenClaw. The two populations want different things and are being served by different products.

Both populations are larger than they were six months ago because OpenClaw proved that the underlying demand was real and large and underserved. The hire did not create the market. It confirmed it.

Sources:

Previously on TheQuery: The Day Claude Code's Moat Disappeared and Cursor 3 Bets on a Different Kind of Developer