Code on Android

🇬🇧 English · 🇫🇷 Version française

A native Gemini coding client for Android — not a wrapper around the official Google app, not a thin webview. A real client that turns your phone into a pocket coding workstation: the model reads and writes files, runs shell commands (compiles, tests, starts servers), generates images, and holds context across hours of conversation. Written in Kotlin + Jetpack Compose.

🎯 What you can actually do with it

• Ask the model to modify a project, not just describe one. It opens files in your workspace (SAF or local folder), edits them literally, and shows you a diff. You approve once — or flip auto-approve and let it iterate on its own.
• Run shell commands from the conversation. The Termux bridge drops the model into your workspace directory: python foo.py, npm test, cargo build, pip install …, curl, git status. Backgrounded processes (servers, watchers) keep running when the model's turn ends.
• Generate images inline. Both Imagen (dedicated picker in Settings → Model) and Gemini 2.5 Flash Image ("Nano Banana", auto-enabled when you pick it in the top-bar dropdown) save their outputs to the chat bubble as thumbnails.
• Send images for the model to analyse. Tap the image icon, pick any photo from the gallery — it's sent as inlineData base64 in the next turn. The model can OCR, describe, or reason about the image.
• Survive long sessions. The app reports live token usage and auto-compresses the conversation into a fresh summary once the context window fills up, so you keep talking without 400 errors.
• Autosave every turn. Close the app, come back three days later, the conversation is exactly where you left it. Name and save snapshots from the drawer for archive.
• Export anywhere. Drawer → Export as Markdown opens the Android share sheet — send the full conversation (text, code blocks, tables, image references) to any app.

✨ Feature breakdown

• Streaming chat over generativelanguage.googleapis.com. API key encrypted locally in EncryptedSharedPreferences. No server in the middle.
• Function calling with 9 built-in tools: read_file, write_file, edit_file, delete_file, list_directory, glob_files, grep, run_shell_command (foreground or background), generate_image (Imagen). The model decides when to call them.
• Safety on destructive tools: every write_file / edit_file / delete_file / shell command shows an approval dialog with arguments and a diff (for edits) before it runs. One-tap "Auto-approve" toggle for trusted sessions.
• Rich markdown rendering: headings, numbered / bullet / task lists, inline + fenced code with a copy button, bold, italic, GFM tables, blockquotes, horizontal rules. Bare https://… URLs and [label](url) links are clickable and open in your browser.
• Top-bar quick pickers: tap the model name to switch models without opening Settings. Tap the workspace folder to "Open folder" (system Files app) or "Change folder" (SAF picker).
• Multimodal input: attach one or more images per turn; thumbnails render in the user bubble and persist across reloads. 15 MB cap per image.
• Dynamic model list: fetched live from /v1beta/models, no hardcoded catalog. Custom model IDs accepted in Settings.
• Diff viewer in the tool-result bubble for edit_file.
• Two languages: EN / FR interface, full parity.

🛠 Prerequisites

• Android 8.0+ (API 26+).
• A Gemini API key (free tier works for chat; image generation requires billing on the associated Google Cloud project): https://aistudio.google.com/app/apikey.
• Optional — Termux from F-Droid or the GitHub releases (⚠️ not the Play Store version, abandoned since 2020) if you want shell command execution.

To build from source:

• Android SDK (API 34+), JDK 17 on JAVA_HOME.

🚀 Installation

Option 1: pre-built APK

Download the latest APK from the Releases page. Both debug-signed and release-signed APKs are published on each tag.

Option 2: build from source

git clone https://github.com/aciderix/gemini-android-app
cd gemini-android-app
./gradlew :app:assembleDebug
adb install app/build/outputs/apk/debug/app-debug.apk

💻 First-run setup

1. Settings → Account: paste your Gemini API key. It's stored encrypted on device.
2. Top-bar folder name → Change folder: pick a workspace under /storage/emulated/0/ (avoid /Android/data/…, unreachable to Termux). The file tools operate relative to this folder.
3. Settings → Termux shell (optional, one-time): follow the 3-step guide to allow run_shell_command. You need Termux installed and termux-setup-storage run once.
4. Top-bar model name: tap it to pick a model. Use gemini-2.5-flash for everyday coding, gemini-2.5-pro for harder reasoning, gemini-2.5-flash-image-preview if you want inline image generation (requires billing).

🧱 Architecture

Multi-module Gradle project:

Module	Role
:app	Compose UI, ViewModels, Activity
:core-bridge	REST Gemini client, tool registry, Termux IPC, SAF, prefs
:domain	Pure types (GeminiMessage, ToolSpec, GeminiEvent…) — no Android
:ui-components	Shared design tokens (theme, colours)

All network traffic flows through RestGeminiCore, which emits GeminiEvents consumed by ChatViewModel. No DI framework, no Room: SharedPreferences + JSON files under filesDir.

⚙️ Advanced configuration

Auto-compression

The model reports usageMetadata.totalTokenCount on every response. Once total / inputTokenLimit crosses the threshold (default 70 %), the conversation is summarised into a fresh session in the background — a non-blocking banner signals the operation.

Tune the threshold in Settings → Auto-compression (50 % → 95 %) or disable it entirely.

Image generation

• Imagen (imagen-3.0-generate-002 by default, picker in Settings → Model) — called via the generate_image function tool when the model decides one is needed.
• Gemini 2.5 Flash Image / Nano Banana — pick it as the chat model, the app automatically enables responseModalities: [TEXT, IMAGE] on every request so the model returns images inline.
• Images are saved to <filesDir>/attachments/ and render as thumbnails in the bubble; chat exports and reloads preserve them.

Both image paths require billing on the Google Cloud project linked to your API key — the Gemini free tier has a quota of 0 for these models.

Background shell commands

run_shell_command accepts a background: true flag so the model can start long-running processes (web servers, file watchers, training loops) without Termux's 12-second IPC timeout killing them. Output lands in $HOME/.gemini-bg/run-<id>.log, tailable on demand.

🧪 Gemini compatibility

Tested with:

• gemini-2.5-pro / gemini-2.5-flash / gemini-2.5-flash-image-preview
• gemini-2.0-flash
• gemini-1.5-pro / gemini-1.5-flash

Gemma models (gemma-2-*, gemma-3-*) appear in the picker but don't support function calling — they'll work for pure chat but the tool stack won't be available.

🤝 Contributing

Contributions welcome:

1. Fork the project.
2. Create a branch: git checkout -b feature/my-feature.
3. Commit with a clear message: git commit -m "Add my feature".
4. Push: git push origin feature/my-feature.
5. Open a Pull Request.

Dev setup

• Language: Kotlin 1.9.24, target JDK 17.
• UI: Compose BOM 2024.06.00 + Material 3.
• Versions centralised in gradle/libs.versions.toml.

Useful commands:

./gradlew :app:assembleDebug        # debug APK
./gradlew :app:lint                 # Android lint
./gradlew :core-bridge:test         # unit tests

📄 License

Distributed under the Apache 2.0 license. See LICENSE for details.

🙏 Credits

• Inspiration: Gemini CLI by Google.
• Model: Gemini API by Google.
• Shell bridge: Termux — an open-source gem.
• UI: Jetpack Compose + Material 3.

Project link: https://github.com/aciderix/gemini-android-app