Voxta v1.8.4: Character Assistant, Interactive Stage Controls & Streaming ARDY
Hello everyone,
v1.8.4 is primarily a stabilization release, with a lot of work going into memory, prompt caching, form reliability, avatar behavior, and platform fixes. But we still managed to sneak in a few notable new toys: a Character Assistant, scriptable on-screen controls, and a much more practical version of ARDY.
🤖 Character Assistant
Characters now have their own creation assistant, similar to the Scenario Assistant.
You can use it to build and iterate on a character conversationally instead of editing every field by hand. It’s designed to make creating richer characters easier while still leaving you in control of the result.
We also made several improvements around character management:
Resume your most recent chat directly from the character page
Edit runtime and setup tabs on protected characters
See VRoid Hub model conditions of use before importing
Prevent accidentally removing the final enabled participant from a chat
🎛️ Interactive Stage Controls
Scripts can now create on-screen controls inside the stage view using chat.setControls.
This includes controls such as sliders and toggles, opening up new ways for scenarios and integrations to expose interactive settings without relying on chat commands or custom UI work.
Support is also included in @voxta/voxta-client, so external clients can receive and display these controls too.
🕺 ARDY Gets Lighter and Faster
ARDY animation generation can now stream its results, letting animations begin arriving without waiting for the entire generation process to finish.
We’ve also quantized the text encoder, significantly reducing its memory requirements and making ARDY more practical on consumer GPUs. It’s still an experimental and demanding system, but this is an important step toward making generated motion more accessible.
🧍 VRM Animation & Presentation
VRM avatars received a substantial animation and presentation pass:
A default idle animation, so avatars feel alive without additional setup
A listening animation while the user is speaking
Emote reactions now reset the idle-fidget timer
The framing camera stays level so characters properly face the viewer
FBX hip height is baked against the rest pose instead of the current live pose
The Animation Playground also gained a timeline for inspecting and working with animations.
🧠 Memory & Prompt Improvements
This release includes several changes aimed at producing more consistent long-running conversations while improving prompt-cache reuse:
Lexical memory matches are weighted by term rarity instead of only term length
Memory merging is now structurally safer
Contradictions are rewritten as history instead of being merged into the current state
Improved summarization and memory prompting
Instruct prompt blocks are ordered by volatility for better prefix-cache reuse
The rendered clock is rounded down to 15-minute intervals instead of invalidating the cache every minute
We also fixed a scripting issue where oneOf would not properly randomize over results returned by matchFiles.
🎭 Stage, Presence & Chat Improvements
Invited guests now appear on the scenario stage
Kicked characters are removed from the stage and other presence views
Chats can be downloaded from the Explore page
System Monitor and Diagnostics behave as full pages when opened outside a chat
KaTeX is retained so LaTeX continues rendering correctly
Markdown tables are styled properly inside the Scenario Assistant
🔧 Modules & Backends
ExLlamaV3 1.3.0
OpenSense 1.4.0
llama.cpp 0.3.45
Claude Code’s resolved executable path is now cached
Linux installations now rely on the system-provided ffmpeg instead of attempting to install it automatically. We also fixed incorrect loading of Semantic Kernel native libraries on Linux.
📱 Voxta Mobile
Audio now uses the proper media-volume channel while chatting and continues behaving correctly after the chat ends.
We also fixed typewriter sounds returning a 500 error on mobile due to secure-path validation.
🛠️ Stability & Fixes
A substantial portion of this release focused on reliability and day-to-day polish:
Fixed multiple form-saving issues
Fixed some modals opening twice, which could prevent changes from being saved
Fixed presets failing to save in some cases
Improved several stage and presence edge cases
Updated llama.cpp to the latest included release
...plus plenty of smaller fixes and refinements throughout Voxta.
Thank you all for the continued support and feedback. Whether you’re building characters with the new assistant, adding interactive controls to a scenario, generating animations with ARDY, or simply enjoying a more stable Voxta, we hope you enjoy v1.8.4.
After subscribing to the required Patreon tier, you can get the latest version at: https://portal.voxta.ai/
🎁 Bonus: ARDY Animation Pack for VRM Avatars
We’ve also attached an Animations.zip pack containing animations generated with ARDY.
To install them:
Extract Animations.zip
Open your character’s Assets tab in Voxta
Drag and drop the included VRM folder into the character’s assets
Voxta will recognize the folder automatically, and your character will begin using the included animations. They’re not perfect, but they provide a good starting set, and a fun demonstration of what ARDY can already generate.