An XDA writer reported this week that they turned their Android phone into a working AI agent using RikkaHub Agent paired with a local LLM. The agent installed apps on its own, controlled entirely via browser, without shipping a single data point to a cloud model. That stack is now within reach for anyone willing to install the right tools. This article covers seven Android apps that run local LLMs or automate the phone for readers who want to build the same setup. Whether you are exploring best apps for local LLM phone automation or looking to reduce cloud dependency, these picks run the full spectrum from simple chat interfaces to sophisticated automation platforms that turn your phone into a true AI agent.
What to look for in a local LLM phone automation app
- On-device inference. The model runs entirely on your phone with no cloud round trip, keeping data local and working offline.
- Support for GGUF or MLC model formats. These are the two practical quantized model formats on Android; GGUF is simpler to use, MLC offers faster inference.
- Function calling or MCP tool support. This is what converts a chat interface into an agent. The LLM must call system functions to read notifications, tap buttons, or send messages.
- Automation hooks. Integration with accessibility services, Android intents, or Tasker plugins lets the LLM trigger actions and control the phone itself.
- Battery and memory profile that survives daily use. Quantized models under 8 billion parameters are realistic on modern mid-range phones; anything larger drains battery fast or requires a flagship device.
Quick comparison
| App | Best for | Platforms | Free plan | Starting price | Rating |
|---|---|---|---|---|---|
| RikkaHub | All-round local LLM client | Android | Yes | Free (open source) | 4.4 |
| PocketPal AI | Fully offline GGUF support | Android, iOS | Yes | Free | 4.2 |
| ChatterUI | Power-user chat and roleplay | Android | Yes | Free (GitHub) | 4.3 |
| MLC Chat | MLC framework inference | Android, iOS | Yes | Free | 4.1 |
| Layla | Mobile AI companion | Android | Yes | Free | 4.2 |
| Tasker | Deep automation platform | Android | Yes | $4.99 one-time | 4.5 |
| MacroDroid | Beginner-friendly automation | Android | Yes, limited | $4.99 one-time | 4.3 |
The apps
1. RikkaHub
RikkaHub is an open source local LLM client built for Android from the ground up. It supports GGUF models, runs inference on-device, and the RikkaHub Agent fork includes full agentic capabilities: the LLM can call functions, control your phone, and chain actions without human prompting. Load a quantized 7B model and watch it navigate your phone autonomously.
Where it falls short: The UI is developer-focused and assumes familiarity with model quantization and GGUF format. Documentation is sparse, and the agent features are newer territory with rough edges on complex instruction chains.
Pricing: Free and open source.
Platforms: Android.
Download: GitHub releases
Bottom line: The most ambitious local LLM phone automation platform available, and exactly what the XDA writer used to build their AI agent.
2. PocketPal AI
PocketPal AI is a fully offline chat interface built around GGUF model support. It downloads models from Hugging Face, runs them entirely on your phone, and does not cache or sync conversations to the cloud. Works on both Android and iOS, making it the cross-platform option if you test on multiple devices.
Where it falls short: Chat-only for now; no automation or function calling yet. If you need your LLM to control your phone, this app talks to the model but does not let the model talk back to Android.
Pricing: Free with no paid tier.
Platforms: Android, iOS.
Download: Google Play
Bottom line: The safest entry point for offline LLM chat on Android if privacy and simplicity matter more than automation.
3. ChatterUI
ChatterUI is a power-user chat client with built-in roleplay and character support. It connects to local GGUF models via llama.cpp backend, supports formatting for creative writing, and includes jailbreak-resistant system prompts. Popular with users building custom AI personas and testing model behavior across different prompt templates.
Where it falls short: Steeper learning curve than PocketPal. No native automation or phone control; intended for chat experiments, not agent behavior.
Pricing: Free (open source, GitHub releases only).
Platforms: Android.
Download: GitHub releases
Bottom line: Ideal if you are experimenting with prompt engineering and want fine-grained control over model behavior and persona design.
4. MLC Chat
MLC Chat is built around the MLC framework, which compiles models to native machine code for faster on-device inference than traditional llama.cpp. Load a model once and it runs at near-native speed. Available on both Android and iOS, with straightforward model management through the app UI.
Where it falls short: Model availability is smaller than GGUF; only models pre-compiled for MLC will run. Setup requires understanding which models are MLC-compatible before you download.
Pricing: Free.
Platforms: Android, iOS.
Download: Google Play
Bottom line: Best choice if inference speed matters and you want the latest optimization research from the MLC framework team.
5. Layla
Layla is a mobile-first AI companion with local model support and a focus on natural conversation. It bridges local LLM inference with a polished chat interface designed for daily use. Sync is optional; run offline by default.
Where it falls short: Less customizable than ChatterUI for roleplay or persona engineering. Function calling and automation are not current features.
Pricing: Free with optional premium features.
Platforms: Android.
Download: Google Play
Bottom line: Best balance of ease and capability if you want a local LLM companion that does not feel rough around the edges.
6. Tasker
Tasker is the gold standard for deep Android automation. Record tasks visually, trigger on sensors or intents, and chain actions across your entire device. It does not run LLMs itself, but pairs perfectly with any local LLM app: connect Tasker to RikkaHub or another client via HTTP, let the LLM decide what to do, and Tasker executes those commands at the Android system level.
Where it falls short: Steep learning curve; not beginner-friendly. Requires understanding Android internals, intents, and automation logic.
Pricing: $4.99 one-time purchase.
Platforms: Android.
Download: Google Play
Bottom line: The most powerful automation platform on Android, and a necessary tool if you want to build a full local LLM agent.
7. MacroDroid
MacroDroid is Tasker for people who do not want to learn Tasker’s scripting language. Drag-and-drop triggers and actions, template library, and cloud sync. Also supports GPT integration if you prefer cloud models, though you can build automations with local LLMs too.
Where it falls short: Less capable than Tasker for complex chains. Cloud dependencies if you use the GPT features; local-only setups lose some convenience features.
Pricing: Free with limited automations; $4.99 one-time for unlimited.
Platforms: Android.
Download: Google Play
Bottom line: Best entry point if you want to automate your phone without learning scripting, and happy to start simple before moving to Tasker.
How to pick the right one
If you want a full local LLM agent that controls your phone right now, start with RikkaHub plus Tasker. RikkaHub handles the LLM and agentic logic; Tasker is the bridge to your phone.
If privacy is your only goal and you do not need automation, PocketPal AI is the fastest path: download a GGUF model, open the app, chat offline. Nothing leaves your phone.
If you are building custom AI personas or testing prompts, ChatterUI gives you the control you need. Pair it with Tasker later if you want to add automation.
If you have a newer phone with strong specs and care about inference speed, MLC Chat will feel noticeably faster than GGUF for real-time chat. The tradeoff is a smaller model library.
If you want all-in-one simplicity without tinkering, Layla as a chat tool plus MacroDroid for automation covers most automation needs without learning curve.
FAQ
Can Android phones really run local LLMs?
Yes. Phones from 2021 onward with Snapdragon 870+ or equivalent can run quantized models from 3 billion to 13 billion parameters. A Pixel 7 or OnePlus 9 handles a 7B model smoothly; flagship devices run larger models. RAM above 6 GB helps; below 4 GB becomes difficult.
What is the best local LLM for Android?
Mistral 7B Instruct (GGUF quantized to Q4) is the sweet spot for Android: fast, capable enough for real tasks, and fits on phones with 6 GB RAM. For agentic work, Nous Hermes 2 7B is popular because it handles function calling well. Smaller models like TinyLlama 1.1B run on anything but are too limited for automation.
Is RikkaHub free?
Yes, RikkaHub is open source under the MIT license. No ads, no cloud account, no feature paywall. You do need to find and quantize your own models or download pre-quantized GGUF files from Hugging Face.
Do I need to root my phone to run local LLMs?
No. All seven apps here run without root. Tasker and MacroDroid do not require root for basic automation, though some advanced system-level actions (changing system settings, modifying other apps) benefit from it. Local LLM inference itself never needs root.
