Claude Code Remote Control

1. The attention bottleneck

Most of the time, Claude Code is faster than you at keyboarding but slower than you at making decisions. For small tasks, this is fine — the task is short enough that you can comfortably stay present, watching each tool call and approving each step. But some tasks are not small. Migrating a 50-file codebase to a new API, backfilling tests across an entire module, or running a large refactor can take an hour or more of continuous execution. Staying glued to the terminal for that long is not a productive use of your time.

Remote control mode changes this equation. It lets Claude Code continue executing — reading files, running commands, implementing and testing — while you step away. You can check back in periodically to review progress, answer questions Claude Code surfaces, and steer if something is going off track. The agent becomes something closer to a junior developer you have briefed and sent off to work independently, rather than a tool that requires your constant supervision.

This is not just a convenience feature. The shift from “I supervise every step” to “I review progress at intervals” is a genuine change in how you interact with the tool. Done well, it multiplies your effective output. Done carelessly, it multiplies the scope of mistakes. Understanding both sides of this is essential before enabling it.

2. How remote control works

When you enable remote control mode, Claude Code opens a persistent session that remains active even after you close your terminal or lock your screen. The session keeps a secure connection to Anthropic’s servers that allows you to reconnect and interact with the running agent from any device — including from the Claude.ai mobile interface.

From the agent’s perspective, remote control mode changes the permission model: it operates with more autonomy, continuing to execute without pausing for approval on each step. It does, however, surface its own blocking questions — cases where it genuinely cannot proceed without your input — as notifications that appear on your phone or in the mobile interface.

From your perspective, remote control looks like a live status feed: you can see what Claude Code is currently doing, read the actions it has taken since you last checked, and send new messages that steer its direction. If you want it to stop, you can interrupt from the mobile interface exactly as you would from the terminal.

3. Enabling remote control

To enable remote control mode for the current session, use the /remote-control command at the Claude Code prompt:

/remote-control

Claude Code will respond with a link or a QR code that you can use to connect from another device. Open claude.ai on your phone, scan the code or follow the link, and you will see the current session.

You can also start a session with remote control enabled from the beginning by passing the flag:

claude --remote-control

Once the session is connected, you can close your laptop or lock your screen. Claude Code will keep working. When it finishes a major step or encounters a question it cannot resolve autonomously, it will push a notification to the connected mobile device.

4. Best task types for remote control

Remote control mode delivers the most value on tasks that share three properties: long duration, clear success criteria, and high verification coverage.

Large-scale refactors are the canonical example: migrating from one state management library to another, updating all callsites of a deprecated API, converting a JavaScript codebase to TypeScript. These tasks have many repetitive steps, each of which can be verified by the test suite, and collectively take longer than a developer wants to sit and supervise.

Test backfills — writing tests for existing untested code — are another excellent fit. The process is mechanical (read a function, write tests for its behavior, run them, fix failures, repeat), the success criterion is unambiguous (test coverage above a threshold, all tests passing), and the risk of introducing wrong behavior is low (tests cannot break production code by existing).

Dependency migrations — updating a major version of a library that has breaking changes — benefit from remote control because they require many small, repetitive changes across many files, and the migration guide gives Claude Code a clear specification to follow.

Tasks that are poor fits for remote control: anything involving UI design decisions, architecture choices, or trade-offs that require product judgment. These tasks require frequent human input, which makes the unattended execution model counterproductive.

5. Checking in effectively from mobile

When you check back into a remote control session on your phone, the most valuable things to look at are:

What is the current state? Read the most recent agent output. Is Claude Code where you expected it to be based on the time elapsed? If a one-hour task has been running for two hours and is only 30% complete, something may be wrong.

What did it do? Review the action log — every file written, every command run. Look for anything unexpected: files being modified that were not in scope, commands that took unusually long, tests that are repeatedly failing and being retried.

What is it stuck on? If Claude Code pushed a notification, read the question carefully. Answer it precisely — a vague answer to a blocking question in a remote session leads to the agent making assumptions and potentially compounding the problem.

Is it on track? If the current approach looks wrong from a higher-level perspective (Claude Code is solving a different problem than you asked it to, or using the wrong pattern), it is better to interrupt now and redirect than to let it continue building on a wrong foundation.

I'm going to enable remote control and step away. Your task: add JSDoc comments to 
all exported functions in src/utils/. For each function, read the implementation 
and write an accurate description with @param and @returns tags. After every 5 files, 
run `npm run lint` to verify the comments are valid. Stop if the linter fails more 
than twice in a row.

/remote-control