Codex on Your Phone: How OpenAI's 2026-05-14 ChatGPT-Mobile Codex Release Actually Works (and What It Means for Regular Users)

"Codex on a phone? Does the AI actually run on the phone? Do I need a Mac? Does the Free plan really get it?" This article walks through the Codex-on-mobile feature OpenAI rolled out as a preview on 2026-05-14, using the OpenAI Codex Changelog, Codex Docs, and the official blog post as primary sources. It's for people who are not technical AI users but want to understand what they're looking at before adopting it. Note that Anthropic already shipped Remote Control for Claude Code in February 2026 — Codex on mobile is OpenAI's response.

Primary sources: OpenAI Codex Changelog (developers.openai.com/codex/changelog), OpenAI's official blog post "Introducing upgrades to Codex," the Codex Docs, and the ChatGPT Release Notes. For desktop Codex features and Claude Code comparisons, see our Codex CLI vs Claude Code benchmark. For the Codex Chrome Extension's per-site approval model, see our Codex Chrome Extension write-up. For the broader browser-plus-AI picture, see our ChatGPT Atlas explainer for non-technical users. For the underlying environment-handover protocol, see our MCP complete guide.

1. What "Codex on mobile" actually is — the 2026-05-14 update

Codex is OpenAI's AI coding agent. A coding agent is "an AI that reads, writes, and runs code when you ask it to." Until this release you ran it on a Mac or Linux box via Codex for Mac or the Codex CLI. The 2026-05-14 mobile update didn't put the AI on your phone — it added a remote-control surface inside the ChatGPT mobile app so you can drive a Mac Codex instance from your phone.

The shift, in plain terms: Codex used to be "an assistant you can only command when you're at your Mac." Now it's "an assistant running on your Mac that you can also talk to from the bus stop." The AI itself runs on the Mac; your phone gets a live view plus the ability to approve, switch model, and send new prompts.

The key phrase is "Codex runs from the connected host" — execution happens on the Mac, not the phone. That's a deliberate design property, not just a limitation.

Why this landed now

[Author opinion] This isn't a one-off feature drop; it's a competitive response. Anthropic shipped Claude Code Remote Control in February 2026, three months earlier. Mobile remote control extends an engineer's working time from "desk hours" to "any waking moment," and both vendors moved quickly to plant a flag here.

The other factor is GPT-5.5 stabilizing. Codex's backing model moved 5.2 → 5.3 → 5.5 over a few months, and 5.5's latency profile makes remote round-trips feel acceptable.

2. What you can actually do from a phone as of May 2026

From the OpenAI Codex changelog:

The four operations in practice:

What you can't do from the phone: edit files directly, manually run git push, get a full IDE, or do large-scale code reviews comfortably. Codex on mobile is optimized for "in-motion status check and approval" — real work still happens at the Mac.

3. Mechanism (1): QR pairing — connecting phone to Mac Codex

QR pairing steps

1. On the Mac, update Codex for Mac to the latest version and launch it
2. From the menu / settings, choose "Connect mobile" — a large QR code appears
3. On the phone, open the ChatGPT app, go to the Codex tab, choose "Connect to Mac," point the camera at the QR code
4. The phone says "Connected" — done. Future sessions auto-reconnect

[Official announcement] Per the changelog, "Codex runs from the connected host" — the QR code is purely about mapping which Mac your phone talks to. Both ends must be logged into the same ChatGPT account, so you can't accidentally pair with someone else's Mac.

Network requirements

After pairing, the phone and Mac exchange messages via OpenAI's servers, not directly. That means you don't have to be on the same Wi-Fi — 4G/5G works fine as long as the Mac is online and Codex is running.

Windows pairing is "coming soon" as of 2026-05-17 with no committed date. See the Windows section below for what to do in the meantime.

4. Mechanism (2): Approve, switch model, send new prompt

Viewing live state

Open the Codex tab in the ChatGPT mobile app and you get every thread running on the Mac. Tap a thread to see its current task, recent output, and any pending approvals — basically "the view you'd have sitting next to the Mac," but on a phone screen.

Approving / rejecting

Codex asks for explicit per-command approval before doing anything that touches the host — file writes, shell commands, git pushes. When one of those fires while you're away, the phone notifies you and you tap Yes or No; the answer goes back to the Mac and either runs or doesn't.

The approval model already existed on desktop Codex; Codex on mobile just lets you make those calls when you're not at the Mac. [Author opinion] We've argued this point repeatedly in Sales Claw's design: the real boundary for handing approval authority to an AI is "can you read the full command on the screen you're using?" If a phone screen can't show the whole command without horizontal scrolling, wait until you're back at the Mac.

Switching models

Model selection is mirrored phone ↔ Mac. Switch from GPT-5.5 to Codex Spark from either end and both reflect the change.

Sending new prompts

You can send follow-ups in existing threads and start brand-new threads from the phone. "Fix this bug," "write a test for X" — type it on the commute and the work has already started by the time you're at the desk. Codex opens the repo on the Mac side, so when you sit down the diff might already be ready.

5. Pricing and supported platforms — why Free / Go users get it

Plan availability

[Official announcement] Per the OpenAI Codex Changelog, "Codex in the ChatGPT mobile app is rolling out on iOS and Android across all plans, including Free and Go, in supported regions." Concretely:

[Unverified] OpenAI hasn't announced when preview ends or whether plan-tier feature splits will appear afterwards. Most ChatGPT previews land general availability in 2–6 months, so a late-2026 GA is a plausible guess; verify against the OpenAI Help Center post-release.

Supported platforms

The control side (phone) and the host side (Mac) have different OS coverage:

6. Getting started — three steps for non-technical users

Step 1: prepare the Mac

1. Update Codex for Mac to the latest version (must include Codex CLI 0.130.0 or newer)
2. Launch it, sign in with your ChatGPT account if you haven't already
3. Open menu / settings → "Connect mobile," a QR code appears
4. Set Mac power options so the system doesn't sleep (display sleep is fine; system sleep is not)

Step 2: prepare the phone

1. Update the ChatGPT mobile app from the App Store / Play Store (1.2026.05 or newer)
2. Sign into the same ChatGPT account as the Mac
3. Tap the "Codex" tab in the bottom nav (shows up on Free and Go too)
4. Tap "Connect to Mac" → grant camera permission

Step 3: pair via QR

1. Point the phone's camera at the QR code on the Mac screen
2. A few seconds later, "Connected" appears on the phone
3. Auto-reconnect is now enabled — no need to re-pair after restarts
4. To unpair, do it from either side's settings

What Windows users can do today

As of 2026-05-17, [Official announcement] Windows pairing is "coming soon" with no committed date. If you want to try Codex on mobile right now and you're Windows-only, your options are:

• Borrow a Mac — even one shared Mac in the office is enough for first evaluation
• Use Codex on the web from your mobile browser (limited functionality)
• Wait for Windows pairing — historically Codex CLI ships Windows weeks-to-months after Mac

[Author opinion] If you're planning enterprise adoption, the realistic move is "one person on one Mac, evaluating now, drafting the internal guidelines." Waiting for Windows before touching it leaves you with no policy when the rollout lands and individuals start using their personal phones uncontrolled.

7. Safety boundaries — when you hand approval authority to a phone

Loss / theft risk

Phones get lost or stolen far more often than PCs. Codex on mobile keeps code and credentials on the Mac, so a stolen phone doesn't leak code directly — but a phone with the ChatGPT app left logged in lets the attacker remote-control your Mac's Codex. Mitigations: turn on screen lock + biometric auth, enable any per-app lock the ChatGPT app supports, and learn the "disconnect this device" flow on the Mac side so you can revoke quickly.

"Just tap approve" risk

Full command text often doesn't fit on a phone screen without horizontal scrolling, which makes "approve without reading" structurally more likely than on a PC. We've repeated this in Sales Claw's design discussions: the quality of approvals is what makes or breaks AI-assisted workflows.

Mac sleep dropping the connection

System sleep on the Mac stops Codex for Mac too, so the phone connection drops. Configure Mac power options so display sleep is on but system sleep is off, or accept that "I tried to approve but the Mac was asleep" is going to happen sometimes. For workflows where the Mac sits idle for hours, this becomes a design constraint.

Network path security

Phone ↔ Mac messages go through OpenAI servers, which is exactly what makes' "same Wi-Fi not required" work — at the cost of trusting OpenAI's infrastructure. OpenAI says the traffic is encrypted; [Unverified] whether it's end-to-end (so even OpenAI can't see the content) isn't explicit in the May 2026 docs. For internal codebases, this is something to clear with InfoSec before broad rollout.

8. Business context and Sales Claw — what "in-motion AI approval" means at work

Engineering impact

Engineers, SREs, and on-call rotations get the most direct change. Concrete examples:

• Overnight incident response goes mobile: PagerDuty / Slack page → phone Codex check → fix suggestion → approve from phone → Mac applies
• Small fixes during commute: see a bug report on the train, send a "fix this typo" prompt, PR is waiting at the desk
• Progress checks during meetings: "did the build pass?" in a few taps without opening the laptop
• Weekend critical bugs: triage and light approvals without sitting down at the Mac

For Sales Claw specifically: "Sales Claw runs overnight, errors page the operator, phone checks state, Codex on the Mac proposes a fix, phone approves, Mac applies" is now a viable loop. The old "the only person who can recover is the one at the Mac" constraint loosens, which is the genuine value.

Non-engineering implications

For non-engineers (sales, marketing, executives), there's little direct use of Codex on mobile itself. What's worth tracking is "engineers' effective working time may expand" and "Anthropic and OpenAI are both serious about the mobile surface now" as industry signals.

Split with other tools

For "use AI from the phone" outside of coding, keep using the right tool for the job:

Where Sales Claw fits

Sales Claw is a locally executed OSS focused on delivering messages through corporate contact forms. Codex on mobile is about "remote-controlling Codex on your Mac from a phone" — different layer, different role.

Sales Claw is a policy-controlled autonomous system that reduces mis-send and policy-violation risk via pre-send automated checks, sales-NG detection, CAPTCHA-detected pause, send-rate limits, and audit-log retention. With Codex on mobile in the picture, a realistic combined flow looks like:

• Overnight run: Sales Claw on the Mac processes outreach forms
• Anomaly alert: Sales Claw detects an error and pages the operator via Slack
• Mobile triage: from the ChatGPT app's Codex tab, ask Codex on the Mac to inspect the log and propose a fix
• Approval: trivial fix → approve from phone, non-trivial → go to the Mac
• Recovery: Mac-side Codex applies the fix and Sales Claw resumes

Pre-rollout checklist (seven items)

1. Get InfoSec sign-off for Codex on mobile (code snippets go through a third-party AI service)
2. Make screen lock + biometric auth mandatory in the device policy
3. Limit approval scope to "small fixes only" in an internal rule
4. Explicitly forbid phone approval for sensitive repos, production deploys, and destructive operations
5. Require post-approval review at the Mac, not on the phone
6. Standardize Mac power settings to "always on" for machines used as Codex hosts
7. Re-review the OpenAI Codex Changelog quarterly, especially for Windows pairing GA

As engineering orgs move AI coding-agent approvals to mobile, run overnight outreach with Sales Claw and let phone approvals only handle exceptional fixes. Local execution, pre-send automated checks, and audit logs hold up the bottom layer of AI sales automation.

無料・MIT ライセンス。インストールせずにライブデモも試せます。

無料でダウンロードライブデモを試すGitHub