Pair your laptop: drive your Mac from your phone over Claude Code Connect

There is a category of small tasks every working developer wants to do without sitting down at the laptop. Running a quick git status on your dev box. Restarting a stuck process. Pulling the latest from a branch. Editing a single line of code. Deploying a tested commit to staging. Reading the last 100 lines of a log file. These are 10-second tasks; they cost 5-15 minutes when you have to walk back to your desk for them.

Claude Code Connect is the surface that lets you do them from your phone. Your laptop runs a small daemon paired with your Ajeris account; the agent forwards commands to the laptop and reports back. Any task Claude Code can do in your terminal becomes a thing your phone can ask for.

This post walks through what the surface is, how the pairing works, what works well and what does not, and what changes about how a developer's day flows.

How pairing works

The pairing flow is intentionally low-friction. Three steps.

# On your laptop:
$ npm install -g @ajeris-ai/cli
$ ajeris connect claude-code
> Enter pair code from your phone: ____

# On your phone (text the agent):
You: pair my laptop with claude code
 
Ajeris: Pair code: 7K2P. Run 'ajeris connect claude-code' on
        your laptop and enter the code. Code expires in 5 minutes.

You enter the code; the daemon authenticates; the agent confirms the pair is live. The whole thing takes under a minute.

The pair token is stored encrypted on your laptop. Revoking the pair is one command (or one text message) and tears down the daemon's authentication immediately.

What "drive your Mac from your phone" looks like

A representative scenario. A developer is at lunch when a deploy goes sideways.

You: ssh into my dev box, check the staging logs for the last
     hour, find any 500s and tell me what's hot
 
Ajeris: On the laptop now. Tailing /var/log/staging-app.log.
        2 distinct 500-level error patterns:
 
        1. POST /api/import (47 hits), TimeoutError. Looks
           like the upload chunk size is hitting the 30-sec
           upstream limit.
        2. GET /api/profile (3 hits), KeyError on user_prefs.
 
        Want me to roll back to the last green commit?
 
You: not yet. tail the import logs for any user data context
 
[Agent runs the command, returns the relevant lines]
 
You: ok roll it back
 
Ajeris: Rolling back staging to a3f8d2 (last green commit).
        Deploy started. ETA 90 seconds.
 
[90 seconds later]
Ajeris: Rolled back. Staging is green. Want to leave a comment
        on the failing PR?
 
You: yes, "rolled back due to chunk-size timeout, ping me when you can take a look"
 
Ajeris: Comment posted on PR #1247.

The whole interaction is six messages. The developer was at lunch. The deploy is fixed. The PR is annotated. The team knows.

This is the surface. It is not a coding-from-the-phone surface (writing a 200-line refactor on a 6-inch screen is not the use case). It is a remote-control-of-your-development-environment surface. Quick reads, quick fixes, deploys, rollbacks, log spelunking.

What works well

A short list of things developers tend to reach for most often.

Git operations. Status, diff, log, pull, branch checkout, commit, push. The whole ergonomics of "wait, what am I working on" via phone is solved.

Process management. Restart a stuck node process. Kill a runaway script. Check what's running.

Log reading. Tail a log, grep for an error, summarize the last 1000 lines. The agent reads the log and tells you what's hot.

Deploys. If your deploy is one CLI command (Vercel, Cloudflare, Railway, kubectl), the agent runs it. Many developers wrap their deploys in make deploy or similar; the agent runs the make target.

Editing single lines. "Update the hardcoded API URL in src/config.ts to point at staging." The agent makes the edit, runs tests if you have a test target, commits if successful.

Shell scripts. Anything a developer would type once into a terminal works.

What does not work well

Three honest gaps.

Long-running interactive tasks. A 90-minute coding session with iterative test runs is the wrong use case. Use the laptop directly; the phone is for the 30-second tasks.

GUI tools. Anything that requires a UI (Photoshop, Figma, Xcode build settings) is not Claude-Code-able. The agent is shell-bound.

Privileged ops without a verification step. Running rm -rf on something that matters; dropping a database table; force-pushing to main. The agent will refuse without a confirmation, and even with confirmation we recommend reaching for the laptop for the highest-blast-radius commands.

How the daemon stays safe

A few notes for the security-curious.

The daemon is a small CLI. Open-source, running locally. You can audit what it does.

Pair tokens are per-device. Pairing your phone with your laptop creates a token specific to that pair. Revoking is per-device.

Commands are mediated. The agent does not have raw shell access; commands flow through Claude Code's tool layer, which has its own approval prompts for destructive operations.

Logs are per-device. What you ran is logged on your laptop, queryable later. The agent's side records the command and the timestamp; what's actually executed is on your machine.

For most professional contexts, this is the right level of containment. For paranoid contexts (production servers with crown-jewel data), additional guardrails are recommended; we have more docs on this.

What changes about your day

The pattern that emerges with developers who use this regularly: the laptop stops being something you go back to for small tasks. Big tasks (real coding, design, sustained writing) still happen at the laptop. Small tasks (everything else) happen from wherever you are. The split is sharp once you adopt it; you stop carrying the laptop for the reasons you used to.

A representative weekday for a developer who pairs:

• At desk: real coding, real reviews, real meetings.
• At lunch: "tail the staging logs for the last hour."
• In the car: "deploy the homepage update Mark pushed."
• At kid pickup: "merge the PR if checks pass."
• At dinner: "check if the cron job ran tonight."

The laptop is on the desk. The work continues.

Try it

Install the CLI. Pair. Run "git status" from your phone. The first time you fix a small thing without sitting down, the system sells itself.