Claude Code Remote Control: Access Your Sessions From Anywhere

Last Tuesday I was out for dinner with my sister when I felt the familiar itch. Somewhere on my desk at home, a claude code remote control session was three hours into a migration I'd kicked off before I left. I kept reaching for my phone between bites of pasta, wanting to check on it the way I used to check the baby monitor when my kids were small.

Is it still breathing? Is it crying? Did it knock the crib rail down?

That was the feeling. A long-running agent, alone in a quiet apartment, chewing through a task I'd set up but couldn't supervise.

Except this time, I actually could check. Claude code remote control meant I pulled up claude.ai/code on my phone, tapped into my session, and there it was: the scrollback, the current tool call, the little blinking cursor waiting for the next instruction. I could see what it had done. I could see what it was about to do. I could even type a quick "pause here until I get back" and hit send, right from the table.

My sister asked what I was smiling at. I told her my code was behaving.

This post is about claude code remote control, the /remote-control command that lets you attach to your local Claude Code sessions from a browser or phone. It's part baby monitor, part walkie-talkie, and it changes how I work when I'm away from my desk. Alex is going to walk you through the setup. I'm going to tell you when it actually matters.

What claude code remote control is (and isn't)

Let's get definitions out of the way before anyone gets confused. There are three Claude Code things in the same neighborhood right now and they all sound similar.

Claude Code remote control is a command you run inside your local Claude Code session. It pairs that session with your Anthropic account and exposes it, securely, at claude.ai/code. Your code still runs on your machine. Your agent still runs on your machine. The only thing moving over the wire is the conversation: prompts you type, output Claude streams back, and permission requests you approve.

Dispatch is the router. It decides which agent or which model should handle an incoming request and hands the work off. Different problem, different tool. If you want more on that, we covered the phone-to-desktop flow with Dispatch separately.

Cowork is the separate desktop application that gives you a full IDE-style Claude Code experience outside the terminal. Also different. Also not what we're talking about here.

Remote control is simpler than either of those. It's a live mirror. When you attach from your phone, you aren't starting a new agent in the cloud, you're looking through a window at the one already running on your laptop. If your laptop lid is closed and the thing is sleeping, remote control can't resurrect it. The laptop has to be awake, the session has to be live, the process has to be running. It's a baby monitor, not a babysitter.

Setup: what Alex actually did the first time

Alex here. Maya asked me to walk through setup because she insists I'm more patient with the boring parts. Fine.

Prerequisites first:

1. Claude Code version 2.4 or newer. Run claude --version to check. If you're older than that, run npm install -g @anthropic-ai/claude-code to update, or whatever package manager you used originally. The command simply doesn't exist before 2.4.
2. A logged-in Anthropic account. If claude /login shows you as signed in, you're good. If not, sign in first.
3. A machine that will stay awake. This is the part people skip. I'll come back to it.

Now the pairing flow. Inside an active Claude Code session, type:

/remote-control

Claude Code will print a pairing URL and a six-character code. Open the URL on whatever device you want to attach from. The browser will ask you to sign in to your Anthropic account, then prompt for that code. Type it. You'll see the session appear in your browser within a second or two, complete with live scrollback.

Pairing is one-time per device. Once a browser is trusted, it shows up in your session list automatically. You can revoke any device from claude.ai/code/devices if you lose a phone or hand a laptop to a family member.

Two gotchas I hit the first time. One: if you're on corporate wifi with strict egress rules, the websocket connection to Anthropic's relay might get killed. You'll see the session connect, then drop after about thirty seconds. The fix is usually an allowlist for *.anthropic.com on port 443. Two: if your local terminal is inside a tmux session inside an SSH session inside a VPN, the pairing still works, but any permission prompt in the browser still fires a local OS notification on the host machine. Quiet that down or you'll have a lot of dock bouncing.

For the deeper terminal configuration, my 50 Claude Code tips post covers the keybindings and shortcuts that make remote attaching feel less clunky, especially when switching between phone and desktop mid-session.

Use case one: monitoring the long jobs

Back to Maya. This is the use case I live in.

I run a lot of what I call "set and forget" agent work. Big refactors. Test suite overhauls. Dependency upgrades where I know Claude will hit fifteen small frictions I can't predict. I'll kick one off at 4 p.m. with a clear plan, then leave the house to do whatever normal humans do on a Tuesday evening.

Before remote control, my options were: stay home and stare at the terminal like it owed me money, or leave and come back to a session that had been stuck on a permission prompt for ninety minutes because Claude wanted to run rm -rf node_modules and politely waited for me to say yes.

Now I just watch from my pocket. Scrollback on my phone, fork on my pasta.

The specific thing that makes this work is the permission prompt flow. When Claude wants to run something risky, it pauses and asks. In the browser, that pause surfaces as a card with "Allow" and "Deny" buttons. I can approve a command with two taps while my sister finishes a story about her dog. I don't need to pull a laptop out of a bag. I don't need to find wifi. Cellular is enough.

What doesn't work yet: I can't initiate a brand new session from the browser. Remote control is attach-only. If my laptop crashes at home, I can't spin up a fresh Claude instance on the phone to fix it. For that, I'd need something like Cowork running in the cloud, which is a different architecture entirely.

Use case two: pair programming without the CLI

Alex again. This one surprised me.

I work with a designer who writes a respectable amount of code but has never gotten comfortable in a terminal. She lives in the browser. Figma, Linear, VSCode web, the usual. Before remote control, our pair sessions were me sharing my screen over Zoom while she pointed at things and I typed. Fine, but one-way.

Now I start a Claude Code session on my machine, hit /remote-control, and send her the pairing URL. She joins from her browser. We see the same scrollback in real time. She can type prompts, I can type prompts, Claude sees a single unified conversation and doesn't know or care which of us is driving.

This is genuinely new. We use it for design reviews where she steers Claude toward the visual outcome she wants while I watch the code Claude is writing and catch the structural issues. She gets a terminal without needing one. I get a second set of eyes. Claude gets clearer feedback than either of us would have produced alone.

One honest caveat: there's no per-user cursor yet. If we both type at the same time, it's chaos. We've settled on a soft rule of "one driver, one observer" and swapping roles explicitly. This is the same etiquette you'd use in a shared Google Doc, just with more consequences if someone fat-fingers a prompt.

For teams doing this more formally, the Anthropic engineering blog has started publishing collaboration patterns that are worth a skim.

Use case three: the overnight batch job check-in

Maya, closing out the use cases.

I run a recurring data cleanup every Thursday night. It's scheduled to start at 11 p.m. and typically finishes around 4 a.m. For the last year I've trusted it, mostly, and woken up to either a green checkmark or a wall of red I'd spend Friday morning untangling.

With remote control, my morning routine changed in a small but real way. While the coffee brews, I open claude.ai/code on my phone, tap into last night's session, scroll up to see where things paused or stumbled, and either send a follow-up prompt or close the session and mark it done. Three minutes, standing at the counter, before I've even sat down at a desk.

It's the same instinct as checking a baby monitor first thing in the morning. You don't need full context. You just need to know the shape of the night. Was it quiet? Were there any cries? Is everything still where you left it?

I've caught two things this way that I would have missed otherwise. One was a silent failure where a dependency upgrade had introduced a warning I wanted to address before Friday's deploy. The other was a case where Claude had correctly paused on an ambiguous prompt and waited, patient as ever, for instruction I could now provide from bed.

Security: the part you can't skip

Before you go paste /remote-control into every session you have, some honest warnings.

Remote control extends the surface area of your local environment. Whatever your local Claude session can do, your paired browser can now trigger. That means file writes, shell commands, git operations, everything. Treat a paired session like handing someone the keys to your dev machine.

A few rules I follow:

• Never pair from a device you don't fully control. Hotel business centers, a friend's laptop, shared kiosks. Don't. Use your own phone or your own browser.
• Revoke devices you don't use. Visit claude.ai/code/devices quarterly and clean house. Old phones, sold laptops, test browsers.
• Be loud about permission prompts. The default behavior is to confirm risky operations. Don't auto-approve everything just because tapping "Allow" on your phone feels frictionless. Anthropic's documentation on Claude Code safety is worth a reread every few months.
• Don't paste secrets into the browser pane. Anything you type goes through Anthropic's relay. It's encrypted in transit. It's still not the place for a production API key. Use your local shell for secrets, use remote control for prompts.

For the broader principles of secure AI agent use, the OWASP AI security guidance is still the best starting point I've found, and it applies cleanly to tools like this.

Closing: go check your crib

I keep coming back to the baby monitor. Not because I think Claude Code is an infant, but because the emotional shape of the thing is the same. You set up a process. You leave the room. You trust it, mostly. You still want a little window to peek through.

What remote control gives you is permission to trust and verify at the same time. You get to leave the house. You get to go to dinner. You get to let the overnight job run. And when the itch hits, when you want to know if the code you kicked off two hours ago is still breathing, you pull up the monitor and look.

Try it this week. Start a small session, something you'd normally babysit, and hit /remote-control. Leave your desk. Walk to the kitchen. Open your phone. See what it feels like to look through the window.

Then tell us what use case we missed. We're still figuring out what this changes. The conversation continues.