Letting My AI Assistant Use My Phone
Quite excited about this - I now have an AI agent that can use my phone.
Quite excited about this - I now have an AI agent that can use my phone. First let’s talk about how to do this safely, then I’ll talk about how to do this.
Safety
My general approach with all AI agent workflows is they must run in their isolated containers - OpenClaw runs in a virtual machine on my NAS and it only has read-only access to things like my personal email.
There are very specific things I give it write access to and those are things that I know I can reverse. For example:
My AI agent has it’s own folder on my cloud drive and I make daily backups of this folder so that if something goes wrong I know I can reverse it
It has it’s own Obsidian Vault (my notes app) so it can interact with my notes but within it’s own sandbox
Similarly I have given my AI assistant it’s own phone. I had an old android phone lying around which has now become my AI assistant’s phone. It was wiped clean before setup and now it only has access to the apps I choose to give it access to.
Phone Access Setup
Here’s how I setup phone access for my OpenClaw instance:
Enable developer options on Android
Enable wireless debugging
Give your agent access to an adb skill and then start wireless pairing (link for reference: https://gadgetguideonline.com/android/android-adb-wireless-debugging/)
Once you’ve done that you can have your agent use your phone!
I have found that Wireless Debugging turns itself off occasionally so I am likely going to switch this over to USB debugging at some point.
The obvious question now is what is this going to be used for?
I have found a couple of interesting use cases here and will write about them over time but the general idea is that phones are a lot more trusted vs browsers. Agents using browsers keep running into captchas and other “are you human?” checks.
By giving my agent it’s own phone I can allow it to run more complex workflows without getting interrupted. It also turns out that phone apps condense information onto a screen quite well and this makes it easier for an AI agent to use.
I will write more about some of these use cases in the future.