I gave a language model hands inside Android — and it built me an app factory

Description and use

An LLM gains shell access via Termux on Android, enabling autonomous app development from prompt to APK installation.

Problem

LLM assistants cannot act in the real world; they only produce text that a human must copy and paste.

Outcome

The system enables fully autonomous Android app development, from prompt to APK installation, using only a phone and an LLM.

Workflow

• 1. User runs vibe-run <project> "<task>"
• 2. Script collects project files and builds a prompt with file contents
• 3. Prompt sent to DeepSeek via VibeService /send
• 4. Wait for response stabilization (hash unchanged)
• 5. Parse response and apply file changes
• 6. Build APK using aapt2 → javac → d8 → apksigner
• 7. If build fails, send error log back to model and retry up to 5 times
• 8. On success, install APK via /install-apk

Evidence

Termux Assistant AI is an Android app that bridges a WebView with DeepSeek Chat and simultaneously gives the assistant direct shell access inside Termux via local HTTP servers.

Source

The key insight: if I have Termux, I have a shell. And if a language model has a shell, it can run arbitrary commands, work with files, compile code, publish releases.

Source

If build fails — sends error log back to the model and retries (up to 5 iterations)

Source