Some changes to make things easier and faster

[siilats]

Changed host to localhost, 0.0.0.0 does not work when clicked. Added .env to gitignore. Installed litert-lm with uv so it gets used in python. Changed model to 4B

[fikrikarim]

Hi. Several things:

1. We intentionally use 0.0.0.0 instead of localhost so it's exposed on the local network. For example, you can open the website from your phone if it's on the same network.
2. I also want to keep the E2B as the default model, so that people with lower-end machine can still run this.
3. Could you explain why we need the litert-lm-api-nightly?

[siilats]

I got a red error about not using the libLiteRtMetalAccelerator. You have in readme that you need to symlink -

"The pip package only ships libLiteRtWebGpuAccelerator.dylib. The C++ build includes native Metal
(libLiteRtMetalAccelerator.dylib) which shows ~6x faster prefill in the C++ binary. However:

That nightly fixes the red errors and looks faster, not 6x.
Maybe there is a way to show a blue hyperlink thats not 0.0.0.0 when the server starts? It gives a bad websocket error thats not clear. You can just print message too, don't type 0.0.0.0 into chrome, the ws will not work.
E4 is a lot better, maybe leave that commented in or .env. Someone did a fork with Qwen to this repo, I have been using mlx-community/Qwen3.5-35B-A3B-4bit its also realtime on the M5 Max, much nicer.

Overall amazing project, I did system prompt of

"You are a friendly AI talking to a 3 year old. She is listening "
"through a microphone and you see her through the camera. "
"She likes to talk about Rapunzel hair. Do not use emoticons or bold * symbols."

and she loves it :)

[fikrikarim]

and she loves it :)

Ohhh that's lovely :)

I personally prefer the E2B model since I'm on M3 Pro and the speed is much better with the E2B model. That's a good point that people should be able to easily use a different model.

Regarding the libLiteRtMetalAccelerator, it's not in README, it's in artifacts docs that I used when developing the repo. It might contain mistakes since it's AI generated. Based on this pip website, the nightly is behind on the litert-lm. The nightly is still on 0.10.0.dev while the litert-lm is on 0.10.1

Ok I see what you're saying that the websocket doesn't work with 0.0.0.0.

[siilats]

Regarding the libLiteRtMetalAccelerator, it's not in README, it's in artifacts docs that I used when developing the repo. It might contain mistakes since it's AI generated. Based on this pip website, the nightly is behind on the litert-lm. The nightly is still on 0.10.0.dev while the litert-lm is on 0.10.1

i think one is cli and the other is python bindings

[siilats]

I added another commit to fix the html issues where on barge it would not speak

[fikrikarim]

Hey we just merged the PR. We now have env variable that can change the model, although we only support Gemma 4 that has audio since our architecture relies on it. We also fixed the barge-in issue on main. Thanks for the contribution!