How I Run Claude Code on My Phone for $29/Month — And Why I Stopped Chasing AI Tools

TL;DR

I use Claude Code from my phone. While watching my kid. While waiting for coffee. Before bed, I kick off a translation task and wake up to find it done. The total cost: $29/month — Claude Code CLI (free), DeepSeek API ($27/month), and a mini PC running 24/7 ($2/month in electricity).

I stopped chasing AI coding tools two months ago. No Cursor, no Copilot, no "let me try this new thing." The time I used to spend evaluating tools now goes into building my product. One tool, used deeply, beats ten tools explored shallowly.

This post isn't a tutorial. It's the philosophy behind my setup — and how you can replicate it in 5 minutes.

The Moment It Clicked

Last Tuesday, I was watching my kid at a playground. My phone buzzed — a GitHub issue notification. A CSS layout bug on mobile. Normally, this would wait until I got home, sat down at my desk, opened my laptop, and fixed it.

Instead, I pulled out my phone, opened a browser tab, and typed into Claude Code:

"Check the mobile layout on the /projects page. The cards are overflowing on 375px screens."

By the time my kid came down the slide, the fix was committed and deployed.

That's when I realized: the real unlock isn't a better AI model. It's removing the constraint of "you must be at your desk to use it."

Why I Quit Cursor

I used Cursor's free tier for a while. When it started asking for $20/month, I looked at what I actually needed: a terminal, file access, and an AI that could read my codebase. Claude Code CLI gave me all of that for free. I didn't need to "try" five other tools to be sure — I just needed one that worked.

The decision took 10 minutes. I've spent zero minutes reconsidering it since.

My Setup: Three Components, $29/Month

Here's what I run:

That's it. No cloud IDE subscription, no SaaS platform, no "pro plan" upsell.

Why DeepSeek?

DeepSeek handles about 80% of my daily coding tasks. It's not Claude Opus, but it doesn't need to be. For routine work — writing functions, fixing bugs, refactoring code — it's solid. I'd rate it 8/10 for coding, as long as you give it clear direction.

Here's my actual API cost breakdown after 2 months:

That's 4+ hours of daily usage. Compare that to Claude's API at $3/$15 per million tokens — the same usage would cost $60-90/month. DeepSeek is roughly 3x cheaper for coding tasks.

Configuration is dead simple. Two environment variables:

export ANTHROPIC_AUTH_TOKEN=your_deepseek_api_key
export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic

No SDK changes, no wrapper libraries. Claude Code CLI speaks the Anthropic protocol, and DeepSeek's API is compatible. It just works.

Where DeepSeek falls short:

• Context window is smaller (~64K vs Claude's 200K) — large codebases need more careful prompting
• Complex reasoning (multi-file refactoring, architecture decisions) is noticeably weaker
• Occasionally hallucinates API signatures — always verify generated code

For complex tasks (architecture decisions, deep debugging), I first discuss the approach in DeepSeek's chat app, confirm the direction, then execute it through Claude Code. This two-step process saves tokens and avoids wasted iterations. (I wrote about my Claude Code + DeepSeek workflow in more detail if you're curious.)

Why a Mini PC?

I have a mini PC that runs 24/7. It consumes about 20W on average — that's 14.4 kWh per month, or roughly $2 in electricity. It's always on, always connected, always ready. No cloud VM fees, no "your instance has been stopped" emails.

The Killer Feature: Claude Code on Your Phone

This is the part that changed everything for me.

I tried various remote desktop and terminal apps to access Claude Code from my phone. Most of them were clunky — tiny text, awkward gestures, poor mobile rendering. Then I found a setup that actually works:

Cockpit — an open-source web-based server management tool. One command to install, one command to start. It gives you a clean web terminal (default port 3001, configurable).

Tailscale — a mesh VPN that makes your home network accessible from anywhere. One command to install, one click to connect.

The full chain:

🔍

Phone → Tailscale → Cockpit (port 3001) → Claude Code CLI

That's the entire infrastructure. Two tools, both one-command installs.

What It Feels Like on Mobile

The experience is surprisingly good. Cockpit's web terminal renders cleanly on a phone screen. Here's the actual workflow:

1. Open your phone browser, navigate to your Cockpit URL
2. Log in — you're greeted with a clean web interface
3. Select a project from your pre-added list (each project is a separate Claude Code session)
4. Start chatting — type prompts like you're sending a message

That's it. No terminal commands to remember, no SSH keys to manage. It's a chat interface for your codebase.

You get:

• Multiple projects — switch between different codebases with a tap
• Persistent sessions — close the browser, come back later, the conversation is still there
• Background tasks — kick off a long-running agent, close the tab, check results later

Real Scenarios from My Week

🔍

Watching my kid at the playground. A CSS bug report comes in. I fix it from my phone in 3 minutes.

Waiting for coffee. I ask Claude Code to research GSC API integration for my blog's SEO pipeline. By the time I sit down at my desk, there's a working prototype.

Before bed. I kick off a translation task for a batch of blog posts. I don't wait for it to finish — I just close the browser. When I wake up, the task is done, the files are updated, and the commit is pushed.

Lunch break. I brainstorm product ideas with Claude Code. Not coding — just thinking out loud, exploring possibilities, validating assumptions. The kind of work that used to require "being in the zone" at my desk.

The Philosophy: One Tool, Deep Mastery

Here's what I believe: the time you spend evaluating AI tools is time you're not spending on your product.

I know developers who've tried Cursor, Copilot, Windsurf, Cline, Aider, Continue, and half a dozen others. They spend weekends configuring, comparing, switching. They have opinions about which one is "best."

I don't have those opinions. I have a product.

This isn't about Claude Code being the "best" tool. It's about the cost of context-switching. Every time you switch tools, you lose:

• The muscle memory you built
• The workflow optimizations you discovered
• The mental model of how the tool thinks

Pick one. Go deep. The compound returns of mastery outweigh the marginal gains of switching.

The $20 vs $29 Question

"But Cursor is only $20/month. Your setup costs $29."

True. But my $29 buys me:

• Mobile access — use Claude Code from anywhere
• No artificial limits — no "you've used your 500 fast requests"
• Full file system access — read any file, run any command
• Extensibility — MCP servers, custom scripts, whatever I want
• No vendor lock-in — it's just a terminal; I can swap the API backend anytime

The extra $9 is the cheapest upgrade I've ever bought.

It's Not Perfect

I want to be honest about the trade-offs:

Long-running agents can hang. When Claude Code spawns an agent that runs for 10+ minutes, it sometimes gets stuck. I've learned to check progress occasionally rather than "fire and forget."

Complex tasks still need a desktop. If I'm doing deep architectural work or debugging a tricky race condition, I want a real keyboard and multiple monitors. The phone is for quick tasks, research, and kicking off background work.

It requires a always-on machine. If your mini PC crashes or loses internet, your workflow stops. I've had this happen twice in two months — both times, a quick SSH restart fixed it.

But these are minor compared to the alternative: only being able to use AI when I'm sitting at my desk.

Two Months In: What I Actually Use It For

I've been running this setup for over 2 months now. Here's what surprised me:

My Top 3 use cases (by frequency):

1. Bug fixes — 40% of my usage. A bug report comes in, I fix it from my phone. Average time: 5 minutes.
2. Research & prototyping — 30%. "Research how to integrate X API, write a proof of concept." I kick these off during lunch and review results later.
3. Writing & documentation — 20%. Blog posts, README files, API docs. Claude Code is surprisingly good at technical writing.

What I expected but didn't happen:

• I thought I'd use it for complex architecture work. I don't — that still happens at my desk with multiple monitors.
• I thought the phone keyboard would be a bottleneck. It isn't — prompts are short and conversational.

What I didn't expect but love:

• The "set it and forget it" pattern. Before bed, I kick off a task. Wake up, check results. This alone has accelerated my project by weeks.
• The reduced context-switching cost. Before, a bug report meant "I'll fix it later." Now I fix it in the moment. Less mental load, faster iteration.

Biggest surprise: The quality of my prompts has improved. Typing on a phone forces me to be concise. Concise prompts get better results. It's a feature, not a bug.

Quick Start: 5 Minutes to Replicate

If you want to try this setup, here's the minimum viable version:

1. Install Claude Code CLI

# Install Node.js if you haven't
curl -fsSL https://fnm.vercel.app/install | bash
fnm install --lts

# Install Claude Code
npm install -g @anthropic-ai/claude-code

2. Get a DeepSeek API Key

Sign up at platform.deepseek.com, create an API key, and set it:

export ANTHROPIC_AUTH_TOKEN=your_deepseek_api_key
export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic

3. Install Cockpit

# Ubuntu/Debian
sudo apt install cockpit
sudo systemctl enable --now cockpit.socket

Access it at https://your-machine-ip:3001. The terminal is built in.

4. Install Tailscale

curl -fsSL https://tailscale.com/install.sh | sh
sudo tailscale up

Install Tailscale on your phone too (App Store / Play Store). Both devices join the same network.

5. Add Your Projects

On your desktop, add the projects you want to access remotely:

# Navigate to your project
cd /path/to/your/project

# Start Claude Code — it will remember this project
claude

Each project you start Claude Code in becomes available in the web interface.

6. Access from Your Phone

Open your phone browser, go to https://your-tailscale-ip:3001, log in, select a project from the list, and start chatting.

That's it. You now have Claude Code in your pocket.

Who Is This For?

• Indie developers who want to squeeze productivity out of fragmented time
• Parents who can't always be at their desk but want to keep projects moving
• Remote workers who want AI access during commutes, coffee breaks, or lunch
• Anyone who's tired of "I'll do it when I get to my computer"

You don't need to be a sysadmin. You don't need a powerful machine. You just need a spare computer, a terminal, and the willingness to stop switching tools.

The Bottom Line

I used to think productivity meant finding the right tool. Now I think it means finding one tool and removing every barrier between you and using it.

Claude Code on my phone removed the biggest barrier: location.

The next time you're waiting in line, watching your kid, or lying in bed — ask yourself: what could I build if I had my AI coding assistant right now?

With this setup, you do.