- Java 47%
- JavaScript 34.7%
- Shell 12.4%
- Python 3.1%
- C++ 1.5%
- Other 1.3%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
|
|
||
| .forgejo/workflows | ||
| apps/android-shell | ||
| archive/roland-2026-10-28 | ||
| docs/preview | ||
| plugins | ||
| presets | ||
| profile | ||
| scripts | ||
| .gitignore | ||
| AGENTS.md | ||
| input | ||
| LICENSE | ||
| README.md | ||
dsh (the DeepSeek Harness) ported to Android, running as a phone-resident assistant: the same
agent loop, a curated phone toolset, and a local model that can think entirely on the device.
This is a developer preview, not a release. It exists so people can try the app and tell us what breaks, long before anything is called stable.
What it does
Ask for a thing; it builds and shows it. These runs were produced on the phone, offline, by a 2–4 B model running on the CPU.
![]() A local run end to end: it looks up the tool it needs ( find), writes the files, serves them and opens the page — no tool list in its context. |
![]() Conway's Game of Life, written and served on the phone: "build me a game". |
![]() The same loop as a conversation — "write a simple Todo web app and open it in my browser": five tool calls, no guessing. |
![]() …and the Todo app it produced, running in the browser. |
![]() Themes are plain JSON and cover the whole interface. You do not have to write one: ask WD and it writes the theme for you. |
![]() One of them applied — Night Glider, if you remember the car. |
The model does not carry a list of every tool in its context. It asks a fuzzy search for the capability it needs and gets back a ready-to-copy example — which is how a 2 B model drives a twelve-action filesystem.
What works
- Local mode — Gemma on the phone's CPU, no network. Files, a web server, the browser.
- API mode — the full harness against a cloud provider you configure.
- Themes — the whole UI is skinnable, and themes ship as plain JSON. A dozen ship with it, and you do not need to write one: ask WD for a theme (or a change to the current one) and it writes the JSON. Easy to share as a file, too.
- Phone tools — screenshots, notifications, Wi-Fi/Bluetooth, scheduling, clipboard, toasts.
- Two independent runtimes, switchable at runtime, so a misbehaving local engine cannot take the app down with it.
Install
Download the APK from the latest release. Android will warn about installing from outside the store — expected for a preview.
Updates are reinstalls. Builds are signed with a key we do not yet keep stable, so installing over an existing app fails; uninstall first. Models and chats live in app storage and go with it.
Requirements: Android 8+ (API 24), arm64. Model downloads are 3–5 GB — use Wi-Fi.
What is rough
- Long turns on a local model can still stop early. The local engine occasionally gives up mid-answer; we are still isolating it and welcome hints. API mode does not show this.
- Small models are small: a 2 B model is cheerful and forgetful, a 4 B one is noticeably better at tool calling.
- Battery and heat are real. Inference is CPU-bound on purpose — GPU drivers on Android are not reliable enough for this yet.
Building
scripts/build-apk.sh builds the APK with the Android SDK's own tools (aapt2, javac, d8,
zipalign, apksigner) — no Gradle. A clean checkout needs two things fetched at build time: the
Android SDK and the prebuilt native libraries, which are published in this repository as the release
native-1. The Node runtime and the dsh tree are downloaded by the app on first run.
Tags build and publish: pushing a v* tag produces an APK and a release. See
.forgejo/workflows/release.yml.
About this tree
This is a preview tree. Development notes, measurements and research material stay private;
AGENTS.md describes how the project is organised and where that material lives.
License
Apache License 2.0. Third-party components keep their own licenses: llama.cpp, Node.js /
libnode and the deepseek-harness packages are Apache-2.0 / MIT.







