Software development is moving fast! The same setup I had six weeks ago, is not the same I have now. In fact, I've been doing software development on my phone!
Just last week I was on vacation at Yellowstone, and I built an app from my phone. I chatted with an agent, it pushed the code to GitHub, and I reviewed the PRs. It was wild!
This post walks through the setup that makes that work. It runs on a server in my house, and I can reach it from anywhere.
Full disclosure, I'm a developer advocate at AWS, and a couple of the tools here are Kiro tools.
The box: a TrueNAS home lab
Years ago I bought an old Intel Xeon server and loaded it with hard drives. It's been my home lab ever since, running TrueNAS. For this setup I created a new Ubuntu VM on it called Kiro Crew and do all my coding there.
I went with Ubuntu 24.04 for the long-term support. 26 was an option, but 24 does everything I need.
You don't need a TrueNAS server for this. I've seen people buy a Mac mini or reuse an old machine, and an EC2 instance or any cloud VM works too. I like keeping it local so everything sits on my own network.
What I installed
After the VM was up, I SSH'd in and installed four things.
- Kiro CLI. Any coding harness that runs on a remote server and lets you authenticate there will work. Most of them do now.
- Kiro Crew. This is the agent I actually talk to. It works like a personal assistant for developers, with sessions, artifacts, memories, skills, and scheduled tasks. OpenClaw, Pi, and OpenCode are other options people run on remote servers.
- The GitHub CLI, so the agent can push code and open PRs.
- Tailscale, so I can reach the server from anywhere.
Connecting from my phone with Tailscale
Tailscale puts the server, my laptop, and my phone on the same private network. Wherever I am, I open Tailscale on my phone, SSH into the box with a mobile SSH app, and port forward the Kiro Crew dashboard. I use port 5477 for mine.
The Kiro Crew docs cover remote setup, and the remote and mobile guide goes through SSH tunnels and phone access step by step. You could use something like CloudFlare tunnels for this as well.
Connecting from my laptop with SSH port forwarding
From my laptop it's a single SSH command with local port forwarding:
ssh -L 5477:localhost:5477 YOUR_USER@YOUR_SERVER
Then I open a new connection window and point it at that port. (I also installed Kiro Crew locally on my Mac Laptop, so I can connect to Kiro Crew locally or to a remote instance)
For security, Kiro Crew wants a token before it lets you in. On the server, run:
kirocrew token
It prints a link with a token in it. Open that, and you're talking to your Crew instance.
Working in Kiro Crew
Sessions live along the side of the dashboard. When I start a new one, I usually turn on trust mode so the agent stops asking for approval after every step. That isn't a good idea for everything. For the personal projects I'm building, I'm comfortable with it.
Then I tell it what to do. In the video I had it pull down a repo I already had. As we go back and forth, it sends me screenshots of what it's building, and it can spin up a dev server so I can look at the running app.
By default it put that project in a scratch folder. Normally I pick the working folder I want before I start. The artifacts tab collects the screenshots and files from all my other projects too, so I can check on things I'm not actively working on.
Memories and skills
Memories are new in Kiro Crew. I told it to always use Vue.js, because I'm a huge Vue fan. I also installed a set of skills, so every new project starts with the same defaults.
Reviewing stacked PRs on GitHub
I always tell the agent to push to GitHub. When I want to be careful, I ask for a stacked PR.
GitHub's stacked PR interface works well with coding agents. It breaks a large update into smaller layers, and you review each one on its own.
For a fun project I'll accept the whole stack. For anything else I go through it layer by layer, make changes, and go back to Kiro Crew with something like "please update the PR, this code isn't what I want." Then we go back and forth until it's right.
The next thing I'd add is CodeRabbit, which reviews PRs automatically before you merge. I don't have it on this project yet.
Building from my phone on vacation
When I was at Yellowstone National Park last week, I wanted to try an ASCII library I'd seen on Hacker News. I asked Kiro Crew to build an app with it, and it created everything.
I did all of it from my phone. Tailscale got me onto my network, I chatted with the agent, it pushed to GitHub, and I checked the PRs. It worked really well.
I can also give it one prompt, turn on trust mode, walk away, and come back half an hour later to see everything it's done.
Models and effort levels
I switch between Claude Opus 5.5 and GPT 5.6 Sol. Opus 5.5 works better for me and what I'm building, so that's my default.
I've also been testing different effort levels, including high. I don't have a clear answer there yet, so your mileage may vary.
Specs and parallel sessions
When I want to be very specific, I use spec-driven development mode. Kiro Crew has a spec builder built in. (You can enable it in the "Discover" apps section)
Most of the time, though, I open three or four sessions and let them run in the background while I do other things.
Scheduled tasks and Discord
I'm trying to do more speaking, so I set up a weekly CFP finder. It runs on a schedule and sends me a message with what it found.
Kiro Crew is also connected to my own Discord channel and posts there. So I can either use Tailscale and port forwarding to open Crew on my phone, or I just open Discord and read the messages as they come in.
What does your setup look like?
This is how I'm building apps right now. A server at home runs the agent, my phone and laptop connect to it, and code reaches me as stacked PRs I can review.
Is this what your future looks like? Have you been running agents remotely, or is everything still happening in your editor? Let me know in the comments.
Top comments (1)
What do you use?