Claude Code Remote Control (2026): Driving My Desktop CLI from My Phone, and Retiring the Telegram Bridge I Open-Sourced

Table of Contents

1. My laptop has been in repair for two weeks, and I leave the house with just a phone
2. What I ran before: a self-built Telegram bridge (which I open-sourced)
3. Setup: from zero to phone-attached
4. Honest scorecard: Remote Control vs my Telegram bridge
5. Limits worth knowing from the official docs
6. What this actually means for a one-person company

#My laptop has been in repair for two weeks, and I leave the house with just a phone

Start with the scene. My laptop is in the shop. In the past, that meant carrying a second laptop and using DeskIn remote desktop to reach the machine at home, operating one screen inside a shrunken picture of another, with occasional black-screen rendering glitches for flavor.

These two weeks I have left the house with nothing but my phone. Open the Claude app, tap the Code tab, and the Claude Code session running on my desktop at home is right there: I send instructions, watch progress, tap through permission prompts, and get a push notification when a long task finishes.

The feature is called Remote Control. This post covers three things: how to set it up, how it compares against the self-hosted alternative I ran for months, and the limits from the official docs you should know upfront.

#What I ran before: a self-built Telegram bridge (which I open-sourced)

I might be one of the more qualified people to make this comparison, because before Remote Control existed I built the whole thing myself, using Telegram as the bridge to drive Claude Code from my phone:

• In May, keeping Claude Code plus the official Telegram MCP plugin alive on Windows meant writing a 200-line toolkit to patch three OS-level gaps. I open-sourced it and wrote up the three papercuts.
• By June the setup had run 60 days without an outage, with the stability designs documented in a follow-up post (Chinese only for now).

It worked, and I used it for months. But self-hosted is self-hosted: plugin config to maintain, OS-behavior gaps to patch yourself, and a full redo on any new machine. That 200-line toolkit is what maintenance cost looks like in physical form.

Discovering Remote Control was anticlimactic: after a VS Code update, Claude popped a notification on its own. I tapped it, followed the prompts, and my phone was attached to my desktop's session. Then I tore the bridge down.

#Setup: from zero to phone-attached

Two prerequisites: the CLI is logged in with a claude.ai account (/login; API keys are not supported), and your phone's Claude app signs into the same account. Pro, Max, Team, and Enterprise all work; a Team/Enterprise Owner must flip the Remote Control toggle in admin settings first.

Step one, on the computer. Pick one:

# A. Already inside a Claude Code session: just run the command
/remote-control My Project

# B. Start a new interactive session with remote access enabled
claude --remote-control "My Project"

# C. Server mode: one process hosting multiple parallel sessions
claude remote-control

Step two, on the phone. Also pick one; there is no pairing step:

1. Press space in the terminal to show a QR code, scan it with the Claude app
2. Copy the session URL from the terminal into any browser
3. Open the Code tab in the Claude app (or claude.ai/code in a browser), find your session in the list, a green dot means it is online, tap in

That is the whole setup. The security model is account binding: only devices signed into the same claude.ai account can see or control your sessions, and traffic is TLS-encrypted through the Anthropic API.

#Honest scorecard: Remote Control vs my Telegram bridge

	Self-built TG bridge	Remote Control
Setup cost	Plugin + OS patches + a toolkit (papercuts surfaced over months, fixed in one morning)	Ten minutes
Maintenance	Patch gaps yourself, redo per machine	None
Viewing terminal output	Bot-forwarded text	Native view with full output and progress
Approving permission prompts	My setup could not; permissions had to be pre-widened	Tap on the phone
Sending photos/files to Claude	Roundabout	Native, lands on the computer
Voice input	Depends on the TG client	Native
Push notifications	Build your own logic	Automatic on task completion and pending decisions
Multi-person / group scenarios	TG groups support this naturally	Bound to a single account

That last row is the bridge's one surviving advantage, and the reason I did not tear everything down: the machine-to-human notification half still runs on my TG bot (monitoring alerts and daily reports, one-way pushes with more than one recipient). What got retired is the human-drives-Claude half, and Remote Control does that far better than my bridge ever did.

#Limits worth knowing from the official docs

Two weeks in, I have little to complain about, but these items from the official documentation are worth knowing upfront:

1. Some commands are local-terminal only: /plugin and /resume. Everyday prompting, model switching (/model), and usage checks work from the phone.
2. No Bypass permission mode from the phone: only Manual, Accept edits, and Plan. Fully-automatic approval still requires the computer.
3. When the computer-side process ends, the session goes offline: within about four hours, claude --resume (or claude remote-control --continue) reattaches it and the phone reconnects on its own; after that, start fresh.
4. Several environment variables block the feature entirely: DISABLE_TELEMETRY, DO_NOT_TRACK, a custom ANTHROPIC_BASE_URL (gateway and proxy users take note). Check these first when you hit a wall.
5. Dialogs auto-close after five minutes without a response; grant the app notification permission, and on Android exempt it from battery optimization.

#What this actually means for a one-person company

The real saving is not the build time. It is mental residency: every day a self-hosted bridge runs, a small background process in your head keeps wondering whether it is still alive. Handing that infrastructure back to the vendor returns all of that attention to product work. That is the true accounting of this upgrade.

My Telegram toolkit is still on GitHub. If your scenario is group notifications or multi-person access, it still earns its keep. If you just want to drive your own Claude Code from wherever you are, use Remote Control and stop building bridges.

New to Claude Code? Start with the hands-on tutorial (Chinese only for now).

---

Feature behavior and limits in this post follow the official Claude Code documentation as of August 2026; the docs do not state a launch date or minimum CLI version, so this post does not guess. The personal-experience sections cover roughly two weeks of daily use starting mid-August 2026.

FAQ

Can you control Claude Code from your phone?

Yes. The official feature is called Remote Control. Run /remote-control inside Claude Code on your computer (or claude remote-control in a terminal), then connect from the Claude mobile app's Code tab, by scanning a QR code, or by opening claude.ai/code. From the phone you can send prompts, watch output, approve permission prompts, attach photos and files, and receive push notifications when long tasks finish.

What plan or setup does Remote Control require?

It works on Pro, Max, Team, and Enterprise (a Team/Enterprise Owner must enable it in admin settings first). The CLI must be logged in with a claude.ai account (API keys are not supported), and the phone signs into the same account; no pairing step. Note that several environment variables block the feature entirely, including DISABLE_TELEMETRY, DO_NOT_TRACK, and a custom ANTHROPIC_BASE_URL.

What happens if the computer-side CLI stops or disconnects?

The phone shows the session as offline. Within about four hours, running claude --resume in the same directory (or claude remote-control --continue for server mode) reattaches the same session and the phone reconnects automatically. If the local network is down for more than about ten minutes, the server-mode process exits and needs a restart.

What can't you do from the phone?

Some commands are local-terminal only (such as /plugin and /resume); permission modes are limited to Manual, Accept edits, and Plan, with no Bypass; and some interactive dialogs only open in the terminal. Dialogs also auto-close after five minutes without a response. Everyday work, prompting, watching output, approving permissions, switching models, all works fine.