穆再排尔·穆合塔尔
← All work

03 2026 / Open source · Android

Nuclear Boy

An AI coding agent that lives on an Android phone.

Role
Founder / independent developer
Period
2025 — 2026
Stack
Kotlin · Jetpack Compose · Hilt · Room · DeepSeek API · SSE
Status
Open source GitHub
Layered translucent panes on black, like sheets of glass or a stack of interfaces.
PL. 03 / Nuclear Boy Layered translucent panes on black, like sheets of glass or a stack of interfaces. 2026

Most coding agents assume a desk: a terminal, a keyboard, a stable connection, a user who can wait. Nuclear Boy assumes a pocket, a touchscreen, a connection that comes and goes, and a user who is walking.

The constraint

The phone is not a smaller laptop. It is a different instrument. It has a camera, a microphone, a battery, and a permission model. Its screen cannot hold a diff; its keyboard cannot hold a function signature. Every product decision follows from those facts.

The first decision was to make the agent visible. A desktop agent can hide its work in a scrollback. A phone agent has to show, in one column, what it is doing, what it is about to do, and what it needs from you. The interface is a single stream, not a chat plus a sidebar plus a settings page.

Architecture

Kotlin and Jetpack Compose, with Hilt for dependencies and Room for local state. The agent talks to DeepSeek over SSE, so tokens arrive as they are produced. A turn has three parts: observe the current state, choose a tool, record the result.

Tools are registered in a registry rather than hard-coded. Skills are longer procedures layered on top of tools. The model can call a skill the same way it calls a tool; the difference is that a skill can contain a sequence of steps and its own instructions.

The hardware controls are the opposite bet: vibration, Wi-Fi, Bluetooth, brightness, clipboard, notifications, twenty-odd capabilities a laptop agent does not have. They make the agent feel native, and they make permission design the real interface. A tool that can turn off Wi-Fi is a trust decision.

The sandbox

The Python sandbox was the hardest part. A phone has no privileged container and no reliable process isolation. The sandbox is narrow on purpose: short programs, text output, an explicit workspace. It is useful for parsing, plotting, and a quick numerical check. It refuses the things a phone should not do.

The tradeoff is capability. A narrow sandbox cannot run a long build or hold a large dependency tree. That is acceptable. The agent is not a replacement for a development machine; it is a way to keep working when the machine is not there.

What I would change

The first version optimized for capability. The second would optimize for interruption. A pocket agent is used in fragments — thirty seconds in a queue, two minutes on a train. It should stop, summarize what it knows, and resume without losing the thread. That is a state-machine problem, not a prompt problem.

The other change is evaluation. I tested tools individually, not their composition under real latency and permission constraints. The next version needs traces from actual use, not a demo script.

An agent that lives where you live has to be smaller, stranger, and more physical than the desktop version.

Measurements

20+ hardware controls
~700ms dual search
1 device
SSE streaming

Next — 04 / 2026

Cold Start

Link prediction when more than half the graph has never been seen.