CapriCore · Capri 1.1.2

A Windows AI desktop companion powered by OpenClaw: bundled local LLMs, Hugging Face GGUF models or OpenAI-compatible APIs, with shared conversation and memory, 51 interface languages, optional voice and explicitly enabled desktop tools.

Website · Discord · Documentation · Roadmap · Contribute

Download the latest Capri for Windows · 1.1.2 release notes · Report a bug

The public repository and releases are CapriChooms/CapriCore. Downloads and update checks do not require a GitHub account, invitation or token.

What Capri can do

• Choose your inference. Use bundled Qwen3.5 models, BETA Gemma/Ministral alternatives, compatible Hugging Face GGUF imports, or saved OpenAI-compatible API connections. Switch from the Local/API control without clearing the conversation.
• Reserve resources. The Local menu has a live CPU/GPU load control, enabled at 85% by default. GPU utilization is an adaptive target; brief peaks remain possible.
• Talk and remember. Text chat, optional local speech recognition and voice, independent language settings, communication styles, searchable memory facts and a local document knowledge base.
• Live on the desktop. Animated expressions and mouth layers, headpats, lightning teleports, window-edge poses, floating dialogue, compact controls, tray mode, themes and resource monitoring.
• Choose how much she moves. Optional desktop roaming, with a 30-second to 30-minute interval and automatic pauses during interaction.
• Use tools with permission. Scoped app/file actions, Spotify, Discord desktop control, web search, deep-thinking mode, recorded workflows and approved user tools. Computer control is BETA and uses a visible indicator with Escape cancellation.
• React to music and calls. Optional local system-audio music detection with BPM animation, and microphone/system-audio call features with a separately configured virtual cable.

Capabilities depend on the selected model, its tool/image support and your permissions. Complex workflows, individual GPU backends and additional voices still need configuration and can require your help; see current limits.

Set up native NPU and editor integration, follow desktop task progress, and check experimental Windows 10 compatibility.

Quick installation

On Windows 10 22H2/11 x64, run the CapriSetup executable for your release, or extract the source package and double-click Install Capri.cmd. The wizard installs private runtimes, selected models and optional speech components. The CPU request classifier is included. Choose local LLMs, Hugging Face model selection, or API mode (BETA). API mode skips local conversation-model downloads and, by default, startup loading. No global Python/Node installation or terminal is required.

For a first installation, complete this wizard before opening Capri.exe. Use the same wizard and installation folder to repair missing runtimes or dependencies.

Upgrading an existing installation. Run the 1.1.2 installer and select your Capri installation folder. This updates the installed runtime and keeps local data, compatible downloaded models and personal settings. The stable update channel uses the public repository.

For voice replies on a supported GPU, choose your Conversation language and the included Capri · Chatterbox voice. Setup installs English or the matching multilingual voice. The CPU Qwen3-TTS option is separate and BETA. No separate voice ZIP, destination folder, recording or training is needed for the included reference. Add or prepare voices later in Settings → Language model → Voice models. Voice installation guide, release packaging.

Capri starts with the microphone and voice replies off on every launch. Their models stay unloaded until you enable the microphone or speaker button. Installing a speech pack does not turn it on automatically at startup; your selected languages and downloaded models are kept.

License/download consent, hardware profiles, language choices, progress, cancellation, repair and selective uninstall are explained in the wizard guide. The installer is unsigned; publisher signing and broader clean-machine qualification remain release work. Capri has read/write access to its own installation folder by default, including code and configuration; revoke it in Settings → File access. Other folders, microphone and screen access require separate grants.

Start from source

For manual local-model development on Windows 10/11 x64, install Node.js 24.22.1 (or a compatible later 24.x), npm and the llama.cpp Windows x64 Vulkan runtime. Set CAPRI_LLAMA_PATH to the full path of llama-server.exe before starting Capri. Open the extracted or cloned folder containing package.json and run:

node --version
npm --version
npm ci
npm run doctor
npm run setup
npm run models:pull
npm run desktop

node --version should show v24.22.1 or a newer 24.x version. npm ci installs the JavaScript libraries from the lockfile; wait for it to finish successfully before continuing. If startup reports Cannot find module or ERR_MODULE_NOT_FOUND, repeat npm ci in this folder. If node or npm is not recognized, install Node.js with npm and reopen PowerShell, or use the graphical installation above. Detailed setup and repair.

Setup explains hardware profiles and saves an ignored local override. The model command downloads the selected text model and memory embedding model. Allow several GB of disk space. Capri manages llama-server locally and can reuse compatible weights from an existing model store. npm stop releases managed resources.

The source package includes the reviewed Capri sprites and animations, plus a procedural fallback avatar. Recordings and trained voice weights remain separate optional packs. Missing speech packs leave text chat available. Installation.

System requirements

Windows 10 22H2 or Windows 11, x64. The smallest supported configuration has 8 GiB system RAM. Requirements above that depend on where the language model runs and which speech components you enable.

RAM, VRAM and free disk space are separate resources. GPU-offloaded weights use VRAM; CPU inference uses RAM, and partial offload uses both. Shared GPU memory borrows system RAM. Leave room for Windows, games and other applications.

API connections (BETA)

Choose Use API during installation, or open Settings → API connections to add, test and switch between saved connections.

For cloud providers, use a key and model ID issued by the chosen provider. For Lemonade or another loopback server, leave the key blank unless the server requires authentication. Tool calls and images must be supported by that exact model; enable their switches accordingly. Cloud APIs need an internet connection and an available provider account; usage charges and provider data policies apply. Messages, relevant memory and tool results are sent to the provider; selected screen images are sent when vision is enabled and used.

API mode skips downloading the local conversation LLM and skips loading it by default. Keep Skip local language model when using an API