Repo Agents

Ayush0Chaudhary/blurr

Panda is an on-device Android AI agent that follows natural language commands and operates the phone UI through the Accessibility Service, built in Kotlin.

  • 988 GitHub stars
  • Kotlin
  • ⚖️ NOASSERTION
  • 🎯 Intermediate
View on GitHub (opens in a new tab) Setup guide
Ayush0Chaudhary/blurr — repo preview

What it is

Panda is a proactive, on-device AI agent for Android. It takes natural language commands and operates the phone's UI by tapping, swiping and typing to complete multi-step tasks across apps. It uses a multi-agent architecture written in Kotlin: the Android Accessibility Service reads the screen and performs gestures, and LLMs handle reasoning and planning. The project is marked work-in-progress.

Who it's for

  • Android developers who want to build or experiment with LLM-driven UI automation
  • Hobbyists and students using it for personal, educational or non-commercial purposes
  • Contributors interested in agentic phone-operator projects

Requirements

Requirements

  • Android Studio (latest version recommended)
  • An Android device or emulator with API level 26+
  • Gemini API keys, or any backend that accepts the documented request payload (GCLOUD_PROXY_URL)
  • Accessibility permission granted to the Panda service
  • API key written in local.properties
  • Personal Use License; commercial use requires a separate license

Setup

  1. Build & run

    Open the project in Android Studio, let Gradle sync all dependencies, then run the app on your selected device or emulator. Write your API key in local.properties; the README says more keys give better speed.

  2. Enable Accessibility Service

    On first run the app prompts for Accessibility permission. Click "Grant Access" and enable the "Panda" service in your phone's settings. This is required for the agent to see and control the screen.

Examples

Backend proxy environment variables

python
python
# the name of these keys donot mean you need google cloud, you can use any servers that can accept requests, i will improve the developer experience in the future by making openapi compatible
GCLOUD_PROXY_URL=<url-of-any-backend-that-accept-responses-like-below-payload>
GCLOUD_PROXY_URL_KEY=<any-password-you-wanna-set-or-leave-empty>

What it does: Points the app at any backend that accepts the documented request payload, so Google Cloud is not required.

Expected request payload

json
json
{
  "modelName": "model-name",
  "messages": [
    {
      "role": "user",
      "parts": [
        {
          "text": "Hello, what can you do?"
        }
      ]
    },
    {
      "role": "model",
      "parts": [
        {
          "text": "I can help you with a variety of tasks. What do you need assistance with today?"
        }
      ]
    }
  ]
}

What it does: The message format your custom backend must accept.

Using Gemini keys directly

bash
bash
GEMINI_API_KEYS=

What it does: Alternative to the proxy: add Gemini keys to play around.

View logs in real time

bash
bash
adb logcat | grep GeminiApi

What it does: Filters device logs to Gemini API activity for debugging.

Pros & cons

Pros

  • Pro:Operates the phone UI through the Android Accessibility Service, so it can perform multi-step tasks across different apps
  • Pro:Backend is flexible: use any server accepting the documented payload, or Gemini keys directly
  • Pro:Free for personal, educational and non-commercial use, including modification and distribution

Cons

  • Con:Project status is work-in-progress and the roadmap is marked as not updated
  • Con:Persistent memory is temporarily disabled
  • Con:Commercial use requires a separate license under the Personal Use License