By empowering a Large Language Model (LLM) with CarToolForge, the Car Tool Playground showcases how an LLM can interact with vehicle functions through a conversational chat interface.
- Free Chat: Engage in an open-ended conversation with the LLM.
- Auto Play: Run a scripted, multi-turn conversation from a pre-configured dataset.
1. Install CarToolForge
Follow the CarToolForge installation guide.
2. Set up a Firebase project
This project uses Firebase AI Logic to make calls to the Gemini API. You'll need to create a Firebase project in the Firebase console and register this app with it.
For detailed instructions, see Step 1: Set up a Firebase project and connect your app.
After configuration, download the generated google-services.json file and place it in the project's app/ directory. If this file is missing, the build will fail with the following error:
Execution failed for task ':app:processDebugGoogleServices'.
> File google-services.json is missing.
Tip
It is highly recommended to build and run the AppFunctionsPilot and firebase-ai sample projects first. This will help you familiarize yourself with the dependencies and potentially save debugging time.
You can build the project using Gradle:
./gradlew clean assembleDebug
The app requires privileged permissions to access vehicle properties and execute function calls. To install it, you'll need a rooted device or emulator.
- Disable Permission Enforcement: First, modify the
build.propfile to disable privileged permission enforcement.
adb root
adb remount
adb shell "sed -i 's/ro.control_privapp_permissions=enforce/ro.control_privapp_permissions=log/g' /vendor/build.prop"
- Install as a Privileged App: Push the APK to the privileged apps directory, such as
/system/priv-app:
adb push app-debug.apk /system/priv-app
adb reboot
- Grant Runtime Permissions: To simplify development (a admittedly bad practice 😜), the app skips runtime permission requests. Instead, using
adb install -gto directly grant all necessary runtime permissions:
adb install -g app-debug.apk
Note
After the initial privileged installation, you can update the agent app using a standard adb install command, provided its permissions in the manifest do not change.
Firebase AI Logic is currently integrated. For a list of available models, see Learn about supported models. The specific model utilized can be changed in FirebaseInference.kt.
The application supports local inference via LiteRT-LM. Push the model file to the application's external storage directory. On Android Automotive (AAOS) device, the path is typically user-specific, for example:
adb push qwen3_4b_q8.litertlm /mnt/user/10/emulated/10/Android/data/org.autoharness.cartoolplayground/files
Select LiteRT-LM as the preferred inference engine and choose the uploaded model from the file selection menu in the settings.
The dataset for Auto Play mode can be found in the test_set.csv. You can edit this file to customize the conversational script.
Take a look at the CONTRIBUTING.md.