Sloppy Joe - Mob running on Android and iOS, modifying a live app with an Agent (looking for testers!)

Short version up top for anyone who just wants the ask:
BEAM has had hot code loading since the jump. Sloppy Joe takes advantage of that by giving the steering wheel of an inherently stable app to an agent which it can reprogram on the fly. I need a few more testers to clear Google’s 14-day / 12-tester gate. If you’ve got an Android phone and some curiosity, opt-in details are at the bottom.

What app do you want? Build it! Here I built a custom weather app that tells me what’s going on.
Here’s about 90 minutes of agentic coding truncated to 12.

What it actually is.
Sloppy Joe is my coding workflow with more guardrails and a second node for stability. These days when I’m building apps with Mob I mostly point Claude at it and telling it to fix screens. I’m interested in putting this workflow into the hands of everyone.

From an architectural perspective, this mostly flowed out of my own workflow. I normally point Claude (or your preferred LLM) at AGENTS.md, then tell it to generate a project with mix mob.new my_world_changing_app. After it generates, attach to it with mix mob.connect and then from there the agent is driving. I see an app that does stuff and I tell the agent what other stuff I want it to do.

So take that idea and bolt on a second BEAM node so this thing can never crash while being modified and put in an MCP server. Now the agent needs to know how to attach to the MCP so we need a proxy to coordinate. Now the connection is a bit unstable and also the phone is doing twice as much work so let’s move the processing part to the server. Now the server needs emulators to have the agents build against. Maybe there will be more than one person at a time that wants to build so now you need a pool. Maybe the user wants to view the app as it’s being built so now you cast the emulator device to the staging website.

So that’s pretty much the whole architecture.

On the models.
Anything that can do MCP will work. Tested with Gemini and Claude. I believe at least one user was using OpenCode.

What I’d want from a tester.
Bare minimum is to install it. Google counts raw numbers but we also need a handful of people that are regularly engaging. In order to engage properly you should know how to hook up an MCP server or Google it. It’s pretty easy. Google actually watches engagement, so a real poke beats a silent install.

The opt-in (the actual ask):

• Reply here or DM me the email tied to your Google account
• I add you to the closed test, you get the opt-in link
• Keep it installed for 14 days, I need 12 in total.

Happy to answer anything about the internals in the thread.

Mob
Sloppy Joe

Any help much appreciated!