An Android app called RikkaHub Agent turns a phone into a local-LLM-powered agent that can write code, run Linux commands, and compile software on its own — as long as you keep saying yes.
A phone running nothing but a local language model just installed and compiled real software on its own: cloning a repository, pulling build dependencies, and building a working voice-transcription tool inside a Linux terminal, on the device, with zero cloud API involved. XDA documented the whole thing using an app called RikkaHub Agent. The interesting part isn't that it happened — it's that the phone asked for permission before every single step, and only acted once it got a yes.
how a phone actually 'installs software on its own'
1
The agent runs a status check and confirms whisper.cpp isn't installed yet.
2
It asks the user for explicit permission before doing anything else.
3
It installs build dependencies inside Termux, Android's Linux terminal environment.
4
It clones the whisper.cpp repository and compiles it from source — about 7 minutes on-device.
5
It downloads a 75MB speech-recognition model so the tool is actually usable.
6
Later, when an audio file needs converting first, it installs ffmpeg on its own without being asked.
That last step is the one worth sitting with. Nobody told the agent it would need ffmpeg. It worked that out from the task in front of it and handled the dependency itself, inside a real Linux environment, on a phone. RikkaHub Agent's tool set goes well beyond installing packages — per its own documentation it can open apps, read and send messages, watch notifications, run scheduled background jobs, and SSH into a remote server, all from natural-language instructions. The model doing the reasoning can live entirely on the phone, or entirely on a separate machine the phone talks to over the network — the agent framework itself is what stays local.
the honest limits
2.6-4.3GB
On-device model size
Gemma 4 E2B/E4B, quantized
~7-8 tok/s
On-device speed
usable for chat, slow for tool-heavy agent work
40-50 tok/s
Networked 27B model speed
Qwen 3.6 27B, running on a Radeon RX 7900 XTX
~32,000 tokens
Prompt size, all tools enabled
up from ~7,000 tokens with tools switched off
That last number explains the real ceiling here. Turning on 80-plus tools doesn't just add capability, it roughly quadruples the context the model has to chew through on every single request, which is exactly why the fully on-device model struggled — a phone-class chip pushing 7-8 tokens per second through a 32,000-token prompt is not a fast experience. XDA's own conclusion after running both setups: genuinely useful for non-frontier tasks like searching files or driving a browser, but it "never will" replace a desktop for running a serious model. That's not a knock on the app. It's just where phone silicon still sits against a 27B-parameter model.
RikkaHub Agent drives real Linux commands inside Termux, Android's terminal environment, rather than a sandboxed simulation. · Unsplash
should you actually try this
Good fit if you already run local models and want a real, functional agent rather than a demo — RikkaHub Agent is free and open-source.
You'll want some comfort with Termux; the interesting capabilities (compiling software, running scripts) route through it.
For anything beyond simple chat and tool calls, plan on a networked model on real hardware — on-device alone is workable but slow.
Not yet a hands-off, set-and-forget agent. Every meaningful action in testing required a manual approval, which is the correct amount of caution for something this new.
Does RikkaHub Agent send anything to the cloud?
Not by design — the model can run entirely on-device or on a separate machine you control over your own network. Neither path routes through a third-party cloud API.
Do I need a powerful phone to run this?
For the agent framework itself, no. For running a capable model fully on-device, yes — XDA's test phone was a Snapdragon 8 Elite flagship with 16GB of RAM, and even then, larger models needed to run on separate, more powerful hardware over the network.
Can the agent do anything without asking first?
In XDA's test, no — every install and file action required an explicit approval. That's this specific app's design choice, not a guarantee every agent framework makes, which is exactly why the permission model matters more than the feature list.
What's the difference between this and a normal LLM chat app on a phone?
A chat app answers questions. RikkaHub Agent can act on your device — installing software, running commands, controlling apps — which is a meaningfully bigger trust decision than picking a chatbot.
This is what agentic AI on the edge actually looks like right now: genuinely capable, gated behind constant permission prompts, and still bottlenecked by phone hardware the moment you ask it to do real work. If you're curious where the ceiling on phone silicon currently sits, it's worth reading how far the other direction goes too — running AI on integrated graphics and even a Raspberry Pi are both further along than most people expect. The phone in your pocket is closer to that end of the spectrum than to a desktop GPU — for now.
