Keep the workspace small
Point the agent at one project at a time. A focused folder gives better results than a whole drive, and it keeps the context tight.
Walkthrough
How to use Kern
From a fresh install to a working AI agent with a real Linux environment, in five steps. The whole thing takes about ten minutes, most of which is downloading Ubuntu.
About 10 minutes
375 MB for Ubuntu
No account needed
Install the APK
Download the build that matches your device. If you are not sure which one to take, the universal APK works everywhere.
| Device | Download |
|---|---|
| Most modern phones | app-arm64-v8a-release.apk |
| x86_64 emulators | app-x86_64-release.apk |
| Older 32-bit phones | app-armeabi-v7a-release.apk |
| x86 emulators | app-x86-release.apk |
| Not sure | app-universal-release.apk |
With the phone connected over USB you can install it directly:
adb install -r app-arm64-v8a-release.apkInstalling by hand works too. Copy the APK to the phone, open it with your file manager, and allow installation from that app when Android asks.
Choose a workspace
A workspace is the folder the agent works in. The Files screen, the editor, the agent and the terminal all share it, so anything the agent changes is immediately visible everywhere else.
- Open an existing folder uses the system picker, so you can point it at a project already on your phone.
- Create an app-managed workspace makes a fresh folder inside Kern's private storage. Simple, and it is always available.
Grant All files access when you can. It lets the terminal open your folder at its real path, so there is one shared tree instead of a mirrored copy. It needs Android 11 or newer; on older versions Kern uses a private mirror and can sync it back on request.
Install the Linux environment
Open the Terminal tab and choose Linux environment, then Install Ubuntu. Kern downloads a pinned Ubuntu 24.04.1 LTS rootfs, verifies its SHA-256, extracts it into a staging folder and only then promotes it into place. A failed download never damages a working install.
Once it is ready you get a real login shell:
root@localhost:/workspace#
apt update && apt install -y python3
[ok] Setting up python3 (3.12.3-0ubuntu2)
This step is optional.
Without it you still get the built-in Android shell, the file explorer, the
editor, the agent and the preview. Ubuntu adds a full userspace with
apt, a compiler and Python.
Connect an AI provider
Go to Settings, then Add provider. Kern needs a base URL, a model id and an API key. It supports anything that speaks the OpenAI chat completions API.
| Service | Base URL |
|---|---|
| OpenAI | https://api.openai.com/v1 |
| OpenRouter | https://openrouter.ai/api/v1 |
| Together | https://api.together.xyz/v1 |
| Groq | https://api.groq.com/openai/v1 |
| Ollama, local | http://127.0.0.1:11434/v1 |
| llama.cpp, local | http://127.0.0.1:8080/v1 |
Press Test first. Kern sends a one-token request, so a wrong key or a wrong model id fails immediately instead of at the first real task. When the test passes, save the profile and pick a model.
Your key stays on the device. It is encrypted with AES-256-GCM using the Android Keystore and is never written to a log. If you use a local model, nothing you type leaves the device at all.
Run your first task
Open the Agent tab and ask for something. Good first prompts:
- "Explain this project and how it is structured."
- "Find every place this reads the API key and list them."
- "Write a README for this project."
- "Run the tests and tell me what failed."
The agent will explore, ask to run commands, and may present a plan before it makes changes. Choose a permission mode first if you want fewer interruptions:
| Mode | Behaviour |
|---|---|
| Ask changes | Every write waits for you. The safest. |
| Auto edit | Edits go through on their own; commands still ask. |
| Full access | Everything runs, except destructive commands, which always ask. |
Good to know
Getting the most out of it.
Use the task list
A multi-step job is far more reliable when the agent has a checklist. It keeps track of what is done and what is left, even across long runs.
Commit before you start
Let the agent work on a clean branch. The tool set includes real Git commands, so a mistake is one revert away.
Watch the preview
For web work, ask the agent to start the server. The preview and its console drawer show exactly what the agent is seeing, so you can correct it early.
Pick a fast model for small jobs
Reading a file does not need a large model. Switch models from the chip above the composer when a task is routine.
Mind the space
The app is about 45 MB. The Ubuntu environment adds roughly 375 MB, and the package cache grows as you install.
Troubleshooting
When something does not work.
Open the Terminal tab, then Linux environment, and install Ubuntu. The built-in Android shell still works in the meantime. If the install stops halfway, use the clean up option on that screen and try again.
Check three things: the base URL ends in
/v1, the model id exists on that provider, and the key is correct. Some gateways only expose the models endpoint to logged-in requests, which is why Kern probes the chat endpoint instead.Grant All files access in Android settings, under Apps, then Kern, then All files access. Without it, folders still work through the system picker and the SAF mirror.
Switch the permission mode above the composer to Auto edit or Full access. Destructive commands always ask, whatever the mode, and that is deliberate.
The download is staged and only promoted when it completes, so an interrupted install never breaks a working environment. Check free space, then retry from the Linux environment screen.
Sessions survive navigation and backgrounding, but not the process being killed. The foreground service keeps them alive while the app is in the background, which covers almost every case.
Ready
Ten minutes to your first agent run.
Download the build for your device and follow the steps above. Everything after that is up to you.