amv007/llm_agent_android repository overview

⁠llm_agent_android

Headless sandbox for running LLM coding agents on Android / Kotlin projects: the agents work without approval prompts inside the container, while the container sees only the project folder it is given.

⁠What's inside

• Ubuntu 24.04, Node.js 22
• Agent CLIs: claude (Claude Code), opencode, codex, GitHub CLI gh
• JDK 17, Android SDK command-line tools, platform-tools, platforms 34, 35, 36, 37.0 with matching build-tools, Gradle, Kotlin compiler
• git, ripgrep, jq, python3, build-essential, tmux, vim, socat

No Android Studio and no emulator - build, lint and unit tests only; adb talks to the adb server of the host.

The image holds no user, keys, logins or agent configs. All of that is mounted at start, so one image serves everyone.

⁠Tags

• latest - the latest build
• X.Y.Z - fixed versions (ru.amv007.version label), for pinning or rolling back

Platform: linux/amd64 only.

⁠Usage

The image is designed to be started by the startdocker_llm launcher from the team scripts repo. The launcher mounts the project and an isolated home, attaches VS Code, bridges adb, and applies the sandbox restrictions:

• no root;
• no host network;
• read-only .git/config and .git/hooks;
• resource limits.

docker pull amv007/llm_agent_android:latest
docker tag amv007/llm_agent_android:latest llm_agents_bld:latest
startdocker_llm ~/prj/my-android-project

Then run claude, opencode or codex in the container shell or in the attached VS Code terminal.

Running it with plain docker run works too, but without the launcher none of the sandbox restrictions are applied.