AetherLAM
An un-sandboxed Large Action Model built to bypass browser l
Hey hackers, I wanted to share the live launch of AetherLAM! Most AI chatbots are locked away inside browser sandboxes. They can generate text, but they can't actually do anything natively on your computer.
AetherLAM uses a decoupled cloud-to-os architecture to break the sandbox. We host a high-speed transactional brain on an always-on PythonAnywhere cloud core to process intents via over-the-air Gemini API routes. It translates unstructured human speech into structured JSON task parameters, then dispatches them down to a native async companion app agent running locally on the user's computer to handle documents or wireless network scans instantly.
Compute tokens decrease ONLY upon verified task extraction success, and free trial accounts get 3 token shards automatically right at launch.
I'd love your brutal product feedback on the decoupled wire loops: https://globalintelpro.com
5 Comments
The decoupled cloud-to-local approach is interesting because it separates intent processing from actually performing the task on the machine. The important test seems to be how reliably that handoff works in real-world conditions without making users think about the architecture underneath.
Yeah you are right!
That’s a good distinction. I’d be curious to hear what you learn once people start using that handoff in real workflows.
You hit the nail on the head, Aryan! That handoff reliability was the absolute hardest part of the solo build.
To make it bulletproof in real workflows without user-facing friction, I completely decoupled the states: the cloud core backend (FastAPI on PythonAnywhere) acts as an always-on 24/7 transaction/intent router, translating prompts via Gemini API into structured JSON directives and pushing them to a disk-backed queue file.
The local machine side uses an asynchronous background tray app with insulated long-polling loops. If the network drops or the cell tower switches while controlling it from a smartphone, the client catches the connection exception, holds the task backlog safely, and runs auto-reboot handshakes within 5 seconds without crashing the OS processes. Real-world handoffs are fully stable now! You can audit the live cockpit grid at globalintelpro 🪐