Hi author here, I've been working day & night over the last week building Jauvex, haven't been so obsessed about building something in a while.
I originally started this project as a two-way voice chat for Claude, because I've been missing that from Codex, which has a great voice chat. I really enjoy talking to my computer and orchestrating agents that way instead of typing.
I tried to improve some things that Codex hasn't yet: Jauvex doesn't talk over you, it wont interrupt you, instead it will wait until you finished talking and only then will reply to you.
I tried to reverse engineer the way that Codex built their voice chat, and I came up with something like this:
1) Immediate response ("One moment, I'm checking that")
2) Context Reinforcement ("So you want to do...")
3) Response Summary ("Here's what I've found about what you've asked...")
Jev is connected optionally as a way to decide whether the first message is a statement, a question, a request, or something else, which helps the voice chat to feel more natural.
I've also built a Pro version which is invite only, and it allows you to add multiple organizations, use the same agents with your team (multiplayer AI agents), routines & workflows, todo lists.
Testing now as well different transcribers and text-to-speech models (testing NVIDIA's over Whisper, but Whisper.cpp has given better results).
And my plan next is to be able to download an agent and move to different workspaces.
The version released currently allows you to experimentally change the provider of the main agent (Jauvex agent) without losing the context, so you can start a conversation with Claude and then continue it with Codex normally.
Hi author here, I've been working day & night over the last week building Jauvex, haven't been so obsessed about building something in a while.
I originally started this project as a two-way voice chat for Claude, because I've been missing that from Codex, which has a great voice chat. I really enjoy talking to my computer and orchestrating agents that way instead of typing.
I tried to improve some things that Codex hasn't yet: Jauvex doesn't talk over you, it wont interrupt you, instead it will wait until you finished talking and only then will reply to you.
I tried to reverse engineer the way that Codex built their voice chat, and I came up with something like this:
1) Immediate response ("One moment, I'm checking that") 2) Context Reinforcement ("So you want to do...") 3) Response Summary ("Here's what I've found about what you've asked...")
Jev is connected optionally as a way to decide whether the first message is a statement, a question, a request, or something else, which helps the voice chat to feel more natural.
I've also built a Pro version which is invite only, and it allows you to add multiple organizations, use the same agents with your team (multiplayer AI agents), routines & workflows, todo lists.
Testing now as well different transcribers and text-to-speech models (testing NVIDIA's over Whisper, but Whisper.cpp has given better results).
And my plan next is to be able to download an agent and move to different workspaces.
The version released currently allows you to experimentally change the provider of the main agent (Jauvex agent) without losing the context, so you can start a conversation with Claude and then continue it with Codex normally.
Next up adding support for Linux & Windows.