DroidClaw: Open-Source Tool Turns Any Old Android Phone Into a Full AI Agent

DroidClaw AI agent controlling an Android phone autonomously

What if you could take that old Android phone sitting in your drawer and turn it into a fully autonomous AI agent — one that reads the screen, thinks about what to do next, taps, types, and navigates across every app on the device?

That is exactly what DroidClaw does. This open-source tool, built with just 10 core source files, connects to any Android device over ADB (Android Debug Bridge) and gives an LLM full control of the phone. No root required. No custom ROMs. Just a USB cable and a goal.

What Is DroidClaw?

DroidClaw is an AI agent framework that controls Android devices through ADB. You give it a goal in natural language — "book the cheapest flight to Delhi on MakeMyTrip" or "send a good morning message to Mom on WhatsApp" — and the agent figures out how to accomplish it.

The agent works by taking screenshots of the device screen, analyzing what it sees using a vision-capable LLM, deciding which action to take (tap, swipe, type, scroll), executing that action via ADB, and then repeating the cycle until the goal is achieved.

It is not controlling one specific app through an API. It is controlling the entire phone the way a human would — by looking at the screen and interacting with whatever is in front of it.

Three Modes of Operation

DroidClaw supports three distinct ways to use it:

Interactive Mode: You type a goal, and the agent pursues it in real-time. You can watch it work, interrupt it, or give it new instructions mid-task.

Workflows (JSON): Pre-defined sequences of actions stored as JSON files. DroidClaw ships with 35 ready-made workflows covering common tasks like checking flight prices, posting to social media, or monitoring stock prices.

Flows (YAML): A more structured approach using YAML configuration files for complex, multi-step automation sequences that run reliably every time.

No APIs Needed — It Uses Your Installed Apps

This is the key differentiator. Most automation tools require API access, which means dealing with authentication tokens, rate limits, paid tiers, and apps that simply do not offer public APIs.

DroidClaw bypasses all of that. Want to check flight prices? It opens MakeMyTrip on the phone and reads the screen. Want to send a Telegram message? It opens Telegram and types it. Want to monitor your investment portfolio? It opens your brokerage app and reads the numbers.

If you can do it by tapping on your phone screen, DroidClaw can do it autonomously.

Turn Old Phones Into Always-On Agents

This is where the project gets genuinely interesting. That old Android phone you replaced last year? Plug it into a Raspberry Pi or an old laptop, connect via ADB, and you have a dedicated AI agent running 24/7.

Some practical use cases the community has explored:

  • Slack standup bot: Automatically posts daily standup updates by reading your calendar and task manager
  • Flight price monitor: Checks prices on travel apps every few hours and alerts you when they drop
  • Telegram digest: Reads through your Telegram groups and sends you a daily summary
  • Social media scheduler: Posts content to Instagram, Twitter, or LinkedIn at scheduled times
  • Stock watchlist: Monitors your portfolio app and alerts you on significant changes

Multiple LLM Providers Supported

DroidClaw is not locked into any single AI provider. It supports:

  • Groq (free tier): Fast inference with free API access — no credit card needed
  • Ollama: Run models locally on your own machine for complete privacy
  • OpenRouter: Access hundreds of models through a single API
  • OpenAI: Use GPT-4o or other OpenAI models
  • AWS Bedrock: Enterprise-grade deployment through Amazon

The Groq free tier option is particularly notable — it means you can run DroidClaw with zero ongoing costs. No cloud subscription, no API fees, just your old phone and a free API key.

22 Actions, 6 Skills, 35 Workflows

Out of the box, DroidClaw comes with 22 built-in actions (tap, swipe, type, scroll, open app, go home, etc.), 6 reusable skills (higher-level capabilities like "navigate to a specific screen" or "fill a form"), and 35 pre-built workflows for common tasks.

The action system is extensible — developers can add custom actions for specialized use cases without modifying the core codebase.

What Works and What Does Not

To its credit, DroidClaw is honest about its limitations. The project documentation clearly states:

  • Complex multi-step tasks may require multiple attempts
  • Some apps with heavy anti-automation measures may resist control
  • Vision accuracy depends on screen resolution and the LLM model used
  • Battery drain is real when running continuously on an old phone

That said, for straightforward automation tasks — the kind where you would otherwise spend 5 minutes tapping through an app every day — DroidClaw works surprisingly well.

The Bottom Line

DroidClaw represents a genuinely novel approach to phone automation. Instead of building integrations with individual app APIs, it treats the phone screen as a universal interface — the same way humans do.

The fact that it runs on old hardware, supports free LLM providers, and ships with just 10 source files makes it remarkably accessible. It is the kind of project that makes you look at that old phone in your drawer and think, "Maybe I should not have shelved you so quickly."

DroidClaw is available on GitHub under an open-source license. Whether you are a developer looking to experiment with AI agents or someone who just wants to automate tedious phone tasks, it is worth a look.