For most developers, the phone has been useful for checking on an AI coding task, approving a permission prompt, or picking up a conversation that started at a desk. Now Claude Code is pushing that idea one step further: you can start a session from the Claude mobile app itself, on a machine that is already running claude remote-control.

It sounds like a small interface change. In practice, it removes one of the more stubborn bits of friction in remote, agent-assisted development: the need to first be sitting at the computer you want to use.

The new flow is simple. Open the Claude app, head to the Code tab, and a machine running Claude Code’s Remote Control service appears as a device card at the top of the screen. Tap the machine, choose a directory, and Claude Code begins a session on that computer. Instead of SSH-ing into a server from a mobile terminal, navigating through directories, launching a session, and then getting back to an AI interface, the phone becomes the front door to a developer’s existing environment.

That distinction matters because this is not Claude Code being run natively on the phone, nor is it simply a cloud-hosted coding agent. The session still executes on the selected local machine, retaining access to its local filesystem, project configuration, installed tools, and MCP servers. The phone is effectively a remote control surface for a machine that already has the right codebase and development setup ready to go.

For developers with a desktop at home, a MacBook at work, a Linux box in the cloud, or a long-running workstation on the network, that turns the Code tab into something closer to a device dashboard. The next task does not have to begin with “Which terminal is this on?” or “Did I leave that SSH tunnel open?” It can begin on the phone, in the few spare minutes between meetings or while away from a desk.

From remote viewing to remote work

Claude Code introduced Remote Control as a way to continue an already running local session from a phone, tablet, or browser. Users could start Claude Code in a terminal, generate a session link or QR code, then monitor and steer that work through the Claude app or claude.ai/code.

That was already useful. AI coding agents are particularly good at jobs that do not need constant attention: tracing a bug across a codebase, writing tests, making a refactor, reviewing a pull request-sized set of changes, or running a long verification process. Remote Control meant a developer could leave their desk while the agent worked, then respond when it needed a decision.

But continuing an existing session and creating a new one are different experiences. The former is a convenience feature. The latter changes where work can start.

A developer heading out for coffee can now decide to investigate a production issue, choose the relevant directory on a remote machine, and begin the Claude Code session without opening a laptop first. Someone reviewing a bug report from a phone can hand Claude the task right away, then return later to inspect its work in the terminal, browser, or mobile app. That is not a replacement for serious code review on a larger screen, but it is a meaningful improvement in how quickly an investigation or background task can get underway.

Anthropic’s update also includes a set of reliability improvements aimed at making the handoff feel less fragile. The company says phone and computer sessions now stay synchronized when a session is resumed on a laptop; a session that exits is shown as offline on the phone within seconds; and connections are designed to recover automatically after a laptop closes briefly or a user changes Wi-Fi networks.

Those details may sound unglamorous, but they are the difference between a demo and a workflow. Developers will not rely on a mobile control surface if it frequently points to stale sessions, loses the live state of a task, or requires a ritual of reconnecting every time a machine sleeps.

Why the local machine still matters

There is a temptation to call every coding agent “cloud coding,” but Claude Code’s Remote Control model is more specific. Anthropic says the mobile and web interfaces are windows into a session that continues to run on the developer’s own machine. Code execution and filesystem access remain local, while messages and session state are synchronized through Anthropic’s service.

That architecture solves a practical problem. Development environments are personal and often messy in ways that are hard to reproduce elsewhere: a local database, a particular version manager, custom scripts, credentials, an internal toolchain, project-specific MCP integrations, and files that are not sitting in a public repository. A cloud agent can be convenient, but it does not automatically have access to that environment.

Remote Control preserves that context. A task started from the phone can work with the same repository, tools, and local paths the developer would have used from their terminal. Anthropic says file-path autocomplete, local MCP servers, project settings, tools, and the filesystem remain available to the remotely controlled session.

It also keeps the role of the phone sensible. Few developers want to write substantial code on a small touchscreen. But plenty want to start a task, provide an instruction, attach a screenshot or file, check progress, answer an approval request, and decide whether an agent should continue. Remote Control is built around those moments rather than pretending the phone should replace the workstation.

The mobile interface can also receive files and images. Photos are passed directly as part of the prompt, while other attached files are downloaded to the connected machine and made available to Claude Code as file references. In a real-world workflow, that could mean snapping a photo of a whiteboard diagram, attaching a crash screenshot from another app, or sharing a short log file while away from the computer that will actually run the investigation.

The mobile agent workflow gets real

The most compelling use of this feature is not “coding from a phone.” It is orchestration.

Imagine a developer who remembers, while commuting, that a dependency upgrade has been waiting for attention. With a Remote Control-enabled workstation online, they could start a Claude Code session in the relevant project folder, ask it to assess the upgrade, identify breaking changes, update the dependency, and run the test suite. By the time they are back at a laptop, the tedious first pass may be complete – with a clear trail of what changed and what still needs human judgment.

The same applies to issue triage. A support report arrives with a stack trace. Rather than bookmarking it for later, a developer can start a session against the correct project directory and ask Claude to locate the likely failure path, reproduce the issue if possible, and propose a fix. The task might still need a human to review the diff, but the delay between noticing a problem and putting an agent on it gets much smaller.

Anthropic has also improved the phone-side experience around the session itself. The company says model and effort-level selections are synchronized between the phone and CLI, heavy sessions open faster on iOS, and several slash commands behave more cleanly on mobile. In particular, /clear resets the mobile view, /compact displays a compaction marker, and /diff opens a native diff sheet.

Again, this is not glamorous product marketing. It is the cleanup work that makes a mobile companion usable after the first novelty wears off. If the phone shows a different model setting than the terminal, or if the state of a long conversation is unclear, it creates uncertainty at precisely the moment a developer needs confidence in what the agent is doing.

There are important boundaries

Starting a Claude Code session from a phone does not mean a phone can wake a switched-off machine or magically reach any computer on the internet. The target machine needs to be online and running claude remote-control, and the developer must be authenticated with an eligible Claude subscription account. Remote Control is available on Pro, Max, Team, and Enterprise plans, though Team and Enterprise deployments require an owner to enable it. API-key-based access is not supported.

The current setup also has limits around infrastructure choices. Anthropic documents that Remote Control is unavailable for certain configurations, including Amazon Bedrock, Google Cloud’s Agent Platform, Microsoft Foundry, enterprise Claude apps gateways, and environments where ANTHROPIC_BASE_URL points to a custom gateway or proxy rather than Anthropic’s API.

Then there is the security tradeoff. Claude Code does not open inbound ports on the local machine for Remote Control; instead, the local session makes outbound HTTPS requests and traffic is protected with TLS and short-lived credentials. That is a better fit for many real-world networks than asking developers to expose a local service to the internet.

Still, the session transcript – including messages, Claude responses, and tool activity – is stored on Anthropic servers to synchronize the experience across devices and support reconnection. The actual execution and filesystem access stay on the local machine, but that transcript storage is an important consideration for companies working with sensitive code or strict data-governance rules. Anthropic explicitly notes that organizations requiring Zero Data Retention cannot enable Remote Control.

For Team and Enterprise customers, Anthropic also offers a beta Trusted Devices option. When enabled by an administrator, a user must enroll the browser, phone, or desktop app and have a recent sign-in before they can view or steer a Remote Control session. Anthropic says biometric checks such as Face ID, Touch ID, Windows Hello, or passkeys are performed locally by the device or browser, rather than sending biometric data to Anthropic.

A subtle shift in AI coding

The bigger story here is not that developers can issue prompts from a phone. Plenty of tools can already do that. The shift is that AI coding agents are increasingly being treated as persistent workers attached to real development environments, rather than one-off chat windows that disappear when a browser tab closes.

Claude Code’s Remote Control setup supports that direction. It can keep local tools and code context in place while moving the conversational control point between a terminal, browser, and phone. Anthropic says connected devices can see live conversation activity, subagents, and workflows, and actions taken from the remote device can affect tasks running on the machine.

That makes the developer’s job less about being physically present at the terminal for every action, and more about deciding what deserves attention, what can be delegated, and when to step back in. The computer remains where the actual work happens. But the moment when that work begins is no longer tied quite so tightly to the computer’s keyboard.