Codex Remote addresses a mundane but important problem with long-running agents: useful work does not always finish while its owner is sitting in front of the host computer. A test can fail after a meeting starts, an agent can reach an approval boundary during a commute, or a browser workflow can need a clarification while the user is in another room.
OpenAI made Codex Remote generally available across all ChatGPT plans on June 25, 2026. The GA release followed a May preview that connected the ChatGPT mobile app to Codex running on a Mac. By GA, users could connect supported Mac or Windows hosts, start or continue work from iOS or Android, review progress, and approve actions from a phone.
Remote does not move an entire engineering environment into the mobile app. The host computer still owns the files, shell, app server, local applications, and project context. The phone is a control surface into a live Codex session. Understanding that boundary explains both the feature’s usefulness and its limitations.
What appears on the mobile side
The remote experience carries more than a chat transcript. OpenAI says it can expose live project context, pending approvals, plugins, screenshots, terminal output, diffs, and test results from the connected environment. A user can answer questions, redirect execution, review findings, move among connected hosts, or start another thread without returning to the desk.
That makes the phone suitable for supervisory actions:
- approve or reject a bounded command;
- choose between two implementation paths;
- clarify a requirement the agent could not infer;
- review a short diff or failed test summary;
- stop a workflow that is taking the wrong direction;
- ask the agent to continue after a known blocker is resolved.
It is less suitable for deep code review, detailed visual QA, or handling a large merge conflict. A small screen can hide context and encourage a quick approval. Remote access increases availability, but it does not lower the amount of evidence needed for a risky decision.
The host still matters
For a local Codex session, the connected computer must remain awake, online, and running the Codex app. If the host loses connectivity, closes the relevant process, or becomes unavailable, the phone cannot manufacture the missing environment. The project does not silently migrate to a mobile cloud workspace.
This model has useful properties. Local tools and files stay where the workflow was configured, and a task can continue using the same terminal, browser state, repository, and approved applications. It also means teams need predictable host power, network, lock, and update policies if they expect unattended work to remain reachable.
OpenAI introduced locked computer use for eligible users so a supported Mac can continue certain remote work after it locks, subject to regional and workspace constraints. That should not be interpreted as permission to disable normal security. Organizations can control remote computer use, and managed workspaces may require explicit admin enablement or role-based access.
From preview to GA
The May preview initially connected ChatGPT mobile to a macOS host. It let users continue threads, inspect live state, and handle approvals. OpenAI expanded Windows remote control later in May, with the Windows machine retaining responsibility for local files and tools.
At GA, Remote Control adopted authenticated one-to-one QR pairing between each supported mobile device and host. Connections used since June 8 remained paired, while older inactive connections needed to pair again. Signing out turns off Remote Control without necessarily deleting the saved pairings.
One-to-one pairing is a meaningful improvement over treating every signed-in device as automatically trusted. It gives the user a visible ceremony connecting a particular phone and computer. Teams should still review the device list, remove old pairings, keep both apps updated, and protect the ChatGPT account with strong authentication.
The GA announcement also mentioned a DigitalOcean Droplet Workspace plugin that can provision a Droplet, configure SSH access, and connect it as a remote workspace. That is a separate plugin capability, not a requirement for Remote and not evidence that every remote task runs on DigitalOcean.
A better pattern for long-running work
Remote control works best when the agent’s task already has explicit success criteria and checkpoints. “Improve this application until it is good” is difficult to supervise from anywhere. “Fix these three failing tests, do not change the public API, and stop before installing a new dependency” gives both the agent and mobile reviewer a clear boundary.
Before walking away from the host, a useful setup includes:
- a named outcome and acceptance tests;
- a clean or understood working tree;
- restrictions on destructive commands and external actions;
- an instruction to pause at ambiguous or billable steps;
- a plan for logs, screenshots, diffs, or test evidence;
- a timeout or stopping condition.
The agent can then make progress independently and surface decisions rather than asking for permission on every harmless read. A phone notification becomes useful when it represents a meaningful boundary, not a stream of routine confirmations.
Security and approval are not the same thing
QR pairing authenticates a remote relationship, but each action still needs appropriate authorization. A trusted phone should not convert a high-impact operation into a one-tap reflex. Review the exact command, destination, affected files, and external system before approval.
Several risks deserve special attention:
- screenshots can expose credentials, customer data, or private messages;
- terminal output may contain tokens or signed URLs;
- a compromised website can attempt prompt injection through the browser;
- a mobile notification can be approved without the surrounding conversation;
- a lost phone may remain paired until the session or device access is revoked.
Use least privilege on the host, avoid placing secrets in prompts, and keep publication, deployment, payment, and destructive changes behind explicit confirmation. Workspace audit and compliance capabilities should be configured according to organizational requirements.
Where Remote saves the most time
Remote is especially useful for work with long passive intervals. Examples include waiting for a test matrix, monitoring a build, reviewing an agent’s repository survey, following a browser-based reproduction, or responding when a multi-step migration discovers an exception.
It can also reduce unnecessary restarts. Without remote access, an agent that reaches a question may sit idle until the user returns. With a clear mobile summary, the user can answer and allow the same thread to continue with its context intact.
The feature is not a substitute for automation. A deterministic CI job should run unattended without a phone. Remote is valuable where judgment remains necessary: the workflow is autonomous enough to make progress but still needs a person at important forks.
Supervising an external media job from a phone
Media operations are a useful example because they combine local assets, external uploads, asynchronous jobs, billable execution, and final visual review. Codex can be configured to call an authorized external MCP service from the host, while Remote lets a user observe the same thread and handle bounded decisions away from the desk.
For example, a Codex thread might follow the Medux video merge tutorial: inspect two approved MP4 files, request upload locations, upload them, create one merge task, store its task ID, poll at a controlled interval, and download the result. The user could remotely approve the external upload or answer a question about clip order.
This is not a native Medux capability inside Codex Remote. Medux remains a separately configured MCP service, and the media job continues in that external system. Remote only exposes the state and controls of the Codex thread coordinating it.
A higher-risk image transformation, such as the Codex MCP head-swap workflow, needs stronger gates. A safe remote sequence would be:
- Codex reports which source and target files it selected, without uploading;
- the mobile reviewer confirms authorization and intended use;
- Codex obtains fresh signed upload URLs and submits one idempotent task;
- the task ID and status remain visible in the thread;
- Codex downloads the result but does not publish it;
- final visual inspection waits for a suitable screen and an accountable reviewer.
Remote access is most valuable here as a continuity layer. It prevents an approved, well-specified task from stalling, while preserving the separation between orchestration, external media processing, and publication authority.
Codex Remote makes agents easier to supervise across time and location. The product’s real achievement is not coding on a phone; it is keeping a human decision-maker reachable without pretending that every decision should be made from a phone.