I want to write code on my phone.
On the train, during a walk, waiting in line. Moments when I can't open my laptop, but my mind drifts to code—that function, that bug, that feature. But typing symbols on a phone is miserable. It's just not practical.
That's what I used to think.
Discovering Coding Agents
Everything changed when I started using coding agents seriously.
Natural language becomes code. Say "refactor this function" and it's refactored. Say "add tests" and tests appear. It can run commands too.
Then it hit me: if I can code with natural language, why not on my phone?
So I Built Pockode
At first, I considered setting up a dev environment in the cloud. But the setup was tedious, the performance mediocre. And honestly, I just wanted to use my familiar home setup.
So I built Pockode.
It lets you control Claude Code on your home PC from your phone. If you have Claude Code installed, one command spins it up. Scan a QR code and you're in.
Pockode is open source. You can disable the official relay and self-host. Your code and runtime stay entirely under your control.
I Stopped Opening Neovim
When I first started building Pockode, I still needed Neovim—at least to check git diffs.
One day, I added git diff and file browsing to Pockode. I was debugging while using Pockode to continue development, and before I knew it—Pockode was all I needed.
Now I only open Neovim when I want to dig deep into the codebase with full-text search.
I once fixed a bug while taking a walk. Gave instructions to the AI on my phone, had it generate code and run tests. By the time I got home, the feature was done.
99% of the Code Was Written by AI
99% of Pockode's code was generated by Claude Code. But I wasn't just delegating blindly—I discussed the design, reviewed the output, and corrected course when needed. Iterate, refine, repeat.
My Go experience barely goes beyond tutorials. I hadn't touched React in ages. Yet I shipped something that works. Go is readable enough that even with fuzzy syntax knowledge, I can tell if the logic is right. The AI suggests the right libraries, proposes approaches I wouldn't have thought of.
If I'd been coding alone, I probably would've abandoned it halfway. AI kept me going.
The Pitfalls of AI (and How to Deal with Them)
AI isn't perfect.
It goes off the rails. Gets stuck in local optimization, loses sight of the original goal. Keeps iterating on a flawed premise. When you notice it, stop it. Explicitly redirect.
It patches instead of fixes. Sometimes it slaps on a band-aid instead of solving the root cause. Patches pile up, and suddenly simple functionality is buried in convoluted code.
It writes meaningless code. This one's insidious. Variables defined but never used. Branches that never execute. Harmless since it doesn't affect behavior. But in the next session, the AI tries to make sense of it: "This variable must exist for a reason." And that's where new misunderstandings begin.
To catch these mistakes, you need to understand the big picture. You can't fully hand things off. But as long as you stay aware, AI is absolutely usable.
Keep Working Away from Your Desk
Coding agents have an annoying problem: permission prompts.
Writing files, running commands—each one pops up a confirmation. Want to step away and let it run? Sorry, it's stuck waiting for approval. But auto-approving everything feels risky.
With Pockode, you can check permission requests from anywhere and approve with a tap. No more waiting.
I used to hate stepping away when I was in the zone. Didn't want to break the flow. Now it's different. Walking around, on the train—I can check on the agent's progress and intervene when needed.
Code from anywhere. That's the life I wanted.
Pockode—give it a try if you're interested.
Original text: https://sijiaoh.com/en/posts/phone-claude-code-and-me/
Top comments (0)