Voxta v1.12.0: Skills, App Assistant & Local Video

Hello everyone,

This release brings Skills: memory books, MCP servers and the built-in assistants can provide instructions that a character loads only when the task needs them, and a new Skills Assistant writes them from your documents. The new App Assistant in the right panel opens pages, finds content, resumes chats, writes characters and changes settings after you confirm. Two new local video modules, MiniMax H3 and LTX-2.5, arrive with a rebuilt video playground. A server-side download queue, private data redactions, VRM face morphs and asset optimization fill out the rest.

Here is what changed.

🧩 Skills

Characters load instructions only when the task needs them, using the Agent Skills SKILL.md format.

• Memory books can contain skills, and chats offer the skills of the books their characters read

• Memory book skills travel in exports and packages

• The Skills Assistant writes and maintains a book's skills and memories from your documents

• A Skills tab in Memory & Skills, and Extract can hand documents to the Skills Assistant

• Memory book entries can point to a skill

• An MCP server's instructions become a skill that loads its tools

• The built-in assistants load their references as skills on demand

• Markdown highlighting for skills and other .md assets

🤖 App Assistant

A new right panel tab that knows the page you are on. It reads freely and edits only after you confirm.

• Reads resources, modules, settings and system status

• Searches your library and the Content Hub

• Finds a chat by what was said in it, and resumes it

• Writes and translates characters, and reads the release notes

• Docked beside editors, and replies show they are still working until done

• Switch the assistant's module from the composer

• The Scenario, Character and Button Style assistants can validate their work, with shorter prompts and new scenario skills for flags

🎬 Video Generation

Two new local video modules, and a rebuilt video playground.

• MiniMax H3: image or reference to video, loops, clip continuation and LoRAs

• LTX-2.5: text or first frame to video with sound, end frames and loops

• Input images can be used as frames or as references

• Hosted video services render several clips at once

• The video playground has start, end, reference and continuation slots, a model and LoRA picker, and a code editor for prompts

• Copy a frame to the other slot, and use the same start and end frame to make a loop

• A Stop button on the preview, and a waiting game while you wait

• Video prompt assistants see the pictures, the clip's length and keyframes

• Video edit can compress to H.264 or AV1 WebM

🎨 Images & Expressions

• The expression generator has its own page and assistant, per-folder prompts, multi-folder selection, a full prompt editor with preview, parallel clips and Ctrl+S to save

• Imagine can start every picture from the reference alone, and supports character sheets as references

• Image viewer: live Animate preview, batch crops, 3:4 and 4:3, and prompt history with up/down

• Image viewer: an Enhance menu for background removal and upscaling

• The image and video playgrounds get a module picker, and Ctrl+C stops a generation

• Local Diffusers supports Anima fine-tunes and kohya Anima LoRAs

• Local Diffusers uses a folder per model type, and LoRAs report what they were made for

• The CivitAI browser can pick a specific file, with its precision and type

• Built-in LoRAs are listed as recommendations while not downloaded

• Model fields show the file name, with a download button when it is missing

🔐 Privacy & Tool Calling

• Private Data Redactions: secrets and custom matches are redacted from everything entering a chat

• Chat augmentations can rewrite tool results before the model reads them

• Tool search: large tool sets (MCP servers by default) wait behind a search

• Tool permissions list every tool as off, ask or allowed

• Allow All (Always) asks for confirmation first

• Workspace and Shell are automatic only in their own chat styles, and opt-in elsewhere

• MCP tool result images reach Claude Code and Codex, and Codex auto-approves Voxta's own MCP tools

• Remote callers to a passwordless account are told to set a password locally

📦 Downloads

• Downloads go through a server-side queue with retries, pause, priority and resume after a restart

• A downloads queue panel behind the downloads pill

• Content Hub downloads resume after the connection drops

🎭 Avatars & Assets

• VRM avatars from VRoid Hub, MMD or VRChat get lip sync, emotions and blinks from their own face morphs (ARKit morphs too), and their legs stand straight

• Optimize assets shows each option's estimated gain, lets you preview and run it, recompresses videos to MP4 or AV1 WebM, and has Run all

• Built-in resources ship WebP and AVIF thumbnails

💬 Chat & Voices

• Multi-turn chat formatting with name prefixes is now the default, and Instruct presets are migrated

• Choices are written in one generation, so they are actually different (requires structured output)

• Interrupted replies keep only the text you actually saw or heard

• Links are spoken as their words or site name, not their full address

• Voices can be filtered by tag groups (Use, Sound, Style, Age)

• Reasoning Effort on OpenRouter and Voxta Cloud has a None choice instead of an Enable Reasoning toggle

• Markdown tables render as tables

• Reopening a chat waits for its previous session to close instead of refusing

• Discord Bot: fewer repeated errors, reactions on passed-over messages, and replies are not lost on stop

🌐 Content Hub & Library

• Choose hub.voxta.ai or hub.voxta.red, and your account decides the rating

• Rating and Style filters, creator profiles, likes, and a link to and from the Library

• The Characters and Scenarios pages load much faster with large libraries

• Galleries, collections and pickers load long lists as you scroll

• Search has an optional details mode that matches descriptions, creators and tags

• Deleting a collection can take its contents with it or un-hide them

• Exported .voxpak file names include what they hold, for example Lily.memorybook.1.0.0.voxpak

• Locked built-ins, like the Voxta docs book, stay as shipped unless creator tools are on

⚙️ Models & Performance

• ExLlamaV3 1.6.0, ONNX Runtime 1.31.0 and updated Python packages

• Speech recognition and a rendering video model are no longer unloaded to make room in VRAM

• More accurate VRAM estimates for Kokoro and ONNX models, for example Parakeet

• The System Monitor splits each service's VRAM by GPU on multi-GPU machines

• Modules interrupt their generation on exit to free the GPU faster

• Server shutdown unloads modules so Python processes exit cleanly, and a crashed Python process says the module stopped and why

• A Shutdown Unused Modules command in quick search

• Add Modules can install several modules at once

🖥️ Voxta UI

• Visual refresh: outline icons, neutral buttons, capsule tags, grey scrollbars and accent-based colors

• The home page is centered on desktop, and notifications use the app's own style

• A Voxta Cloud menu on the credits in the left panel

• Image and video generation get their own menus in the right panel

• Mic and speaker: a click toggles, the chevron opens the menu

• Ctrl or middle-click in the preset picker opens the editor in a new window

• An Edit button beside the Quick Chat Default Scenario in Character Edit

• A copy button next to a newly created API key

• Dropdowns close on an item, not on a misclick

🪟 Voxta Desktop

• The server gets two minutes to unload its modules before being killed on exit

📱 Voxta Mobile

• The Workspace module ships on Android

• The Android open tool is now open_link

🆓 Voxta Lite

New to Voxta? Lite is free and runs on Voxta Cloud, so you can create an account and start chatting without any local setup. If you like it, a subscription gets you the full version, with a large library of integrations, custom modules, and the option to run everything on your own hardware.

This release also carries a long list of fixes, including Voxta Cloud losing earlier words after a tool call, reasoning tokens being reserved twice on OpenRouter and Voxta Cloud, and several image, voice and video modules after Python package updates.

Thank you all, as always, for the continued support and feedback. Whether you are turning your documents into skills, asking the App Assistant to find that one chat, or rendering your first clip with LTX-2.5 or MiniMax H3, we hope you enjoy this release.

Voxta is available to subscribers at: https://portal.voxta.ai/

What's new in Voxta 1.12: Skills, Voxta Assistant, expressions and video (YouTube)