CapriChooms/CapriCore

What it is

CapriCore packages Capri, a Windows desktop character companion built on OpenClaw. It combines a chat assistant with an animated desktop persona and lets the conversation run against bundled local Qwen3.5 models, imported Hugging Face GGUF models, a local OpenAI-compatible server such as Lemonade, or a cloud API (currently BETA). Conversations, memory and an optional local document knowledge base are shared across modes, so switching inference backends does not clear the chat.

The companion also offers optional local speech recognition and text-to-speech, communication styles, animated poses and expressions, tray mode, and permission-gated desktop actions for apps, files, Spotify, Discord and web search. Tool and image support depends on the selected model and the permissions you grant; computer control is explicitly marked BETA.

Install on Windows

Target platform is Windows 10 22H2 or Windows 11 x64. A first install goes through a wizard that sets up per-installation Node/npm/Ollama runtimes and the models you select; no global Node or Python installation is required.

1. Download and run CapriSetup-1.0.4.exe from the v1.0.4 release, or extract the source package and run Install Capri.cmd.
2. In the wizard, choose local models, Hugging Face model selection, or API mode. API mode skips the local conversation-model download and does not load one at startup by default.
3. Finish the wizard before launching Capri.exe. The same wizard can repair missing runtimes or dependencies using the existing install folder.

Users upgrading from the earlier private channel should run the 1.0.4 installer over the existing folder instead of uninstalling first; existing local data, downloaded models and settings are preserved, and update checks move to the public repository. Installation and repair details are in the wizard guide.

Run from source

Manual development on Windows needs Node.js 24.22.1 (or a later 24.x), npm and Ollama, with PowerShell reopened after installing them. From the folder containing package.json:

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

node --version should report v24.22.1 or a newer 24.x. npm ci installs libraries from the lockfile; if startup reports Cannot find module or ERR_MODULE_NOT_FOUND, repeat npm ci in the same folder. npm run setup records hardware profiles without downloading models, and npm run models:pull downloads only the selected text model and memory embedding model. Expect several GB of disk use, and run npm stop to release managed resources. Further instructions are in the setup and repair guide.

Hardware and inference options

8 GiB of system RAM is the floor. The bundled Qwen3.5 models range from 0.8B to 9B, with Ministral 3 3B and Gemma 4 E2B as BETA alternatives; the project publishes RAM/VRAM planning tables for 32K context, and separate estimates for optional speech components. GPU offload depends on the installed Ollama backend and drivers, while CPU inference and partial offload consume more RAM and run slower. A 16 GiB RAM plus 16 GiB VRAM machine is not guaranteed to run the 9B model comfortably once Windows, speech and other applications are counted.

For cloud APIs, choose a preset or a custom OpenAI-compatible endpoint under Settings → External AI provider. Presets include OpenRouter, DeepSeek, Google Gemini, Hugging Face Inference Providers and a local Lemonade server; supply a provider-issued key and model ID, leave the key blank for loopback servers unless authentication is required. Messages, relevant memory, enabled tool results and, when vision is active, selected screen images are sent to the provider, and usage charges and data policies are the provider's.

Hugging Face support accepts GGUF quantizations that the managed Ollama runtime can run; arbitrary Transformers or Safetensors repositories are not automatically compatible, and vision needs matching model and projector files.

Constraints and privacy

• The installer is unsigned; publisher signing and broader clean-machine qualification are still listed as release work. Windows SmartScreen may warn.
• Voice is off at startup on every launch. GPU Chatterbox needs NVIDIA CUDA, or experimental support for selected AMD ROCm GPUs on Windows 11; the CPU Qwen3-TTS option (BETA) adds an estimated 4–8 GiB RAM and may run slower than real time.
• Capri has read/write access to its own installation folder, including code and configuration, by default; other folders, microphone and screen access require separate grants.
• Conversation history, attachments and ordinary memory are stored locally without full encryption; the separately password-protected private memory has its own protection. API credentials are encrypted for the Windows account. Diagnostic exports are saved locally unless you upload them.
• The API text-only clean install is estimated at roughly 3 GiB of downloads and 8 GiB free disk space before local models or speech packs.
• Project code is GPL-3.0-only; character artwork, voices, model weights and third-party packages carry separate terms. Camera support is not implemented.

HotRepo's take

HotRepo recorded CapriCore on the new JavaScript list on 2026-11-14, with 103 stars and 9 forks and a one-hour window in which it gained 3 stars; no all-language daily rank has been recorded. That is a very thin signal and says little about adoption. The technical picture is more informative: the README is unusually explicit about RAM/VRAM budgeting, model compatibility limits, the unsigned installer, and exactly which data goes to cloud providers. The main thing to watch is the distribution path — a Windows installer and prebuilt executables whose relation to the published source is not spelled out, so anyone evaluating it should weigh the documented constraints and the unsigned build before committing a machine to it.