AGENTCODI: Bringing the Codex workflow directly to Android

Hello everyone 👋,

I’d like to introduce a project I’ve been working on over the past few weeks, and also share some of the problems I ran into while trying to make this kind of workflow work properly on Android.

AGENTCODI brings the Codex workflow directly to Android, without requiring Termux, a gateway that has to be started from a console, or a separately configured development environment.

The current public release is 0.4.5 Early Access and runs on Android 10+ ARM64 devices.

It supports signing in with ChatGPT or an OpenAI API key, starting new Codex threads or continuing existing ones, streamed responses and tool activity, approval requests, dynamically loaded models and reasoning levels, and a dedicated project workspace. Files, generated images, and complete workspaces can also be exported directly through Android.

The application and test code itself is written in Java and C++.

One of the more interesting problems came up when I started working on development tools inside the workspace.

For quite a while I was trying to keep Node.js completely out of future AGENTCODI builds. My original idea was to let users install tools such as Node.js themselves through the integrated terminal whenever they actually needed them.

Android had other plans.

Starting with Android 10, an app can no longer simply download a binary into its writable app directory and execute it from there. That made the approach I had planned impractical.

So for upcoming builds I changed direction and started packaging the required Node.js environment directly with AGENTCODI.

The first tests are now working. The terminal is integrated, Node.js is available to the model, and bundled tools can be executed from inside the workspace. This work is still part of development and is separate from the currently released 0.4.5 build.

The next question I’m working on is how to protect that bundled environment against manipulation without getting in the way of the development workflow itself.

That has probably been one of the more interesting parts of building this so far. A lot of the work is less about getting Codex onto a phone and more about figuring out how far you can push this kind of workflow while still living inside Android’s application and security model.

The source code is currently not public, and I’m still unsure whether I want to open it later.

I’m actually curious what people here think about that. If you were building something like this, would you open the source at some point, keep it closed, or wait until the architecture and security model are more settled?

GitHub: https://github.com/Mcpasi/AGENTCODI