ShadowAI — a local-first AI workstation for desktop.

ShadowAI is designed to run locally on your device with minimal reliance on cloud services. To make that possible, it bundles the runtimes and tools needed for local AI workflows, including: • Ollama for local LLM serving • llama.cpp binaries (CUDA, Vulkan, and AMD HIP builds) for GPU-accelerated inference • Voice synthesis (TTS/voice cloning) and speech recognition (Whisper) runtimes • Image, video, and music generation components and their Python runtimes • Document processing tools such as OCR, PDF utilities, and LibreOffice • GPU acceleration libraries for NVIDIA and AMD hardware Ollama models themselves are not bundled by default and can be downloaded separately from within the app. The larger download size comes from bundling the local inference runtimes, media generation pipelines, and multi-backend acceleration libraries needed for a fully self-contained, offline-capable experience.

Yes, to a degree. The biggest contributors are the media generation runtimes — audio-runtime (TTS, voice cloning, Whisper, music) and vision-runtime (image and video generation). Slimmer builds without them are possible and already supported via optional component flags. GPU backend binaries are distributed separately for AMD and NVIDIA, so you only need one. The core app — chat, agents, model management, and Ollama — is significantly smaller without the media runtimes. Fully modular installation is on the roadmap. For now, all packages include the full runtimes.

Yes. ShadowAI supports cloud providers through API keys. You can connect OpenAI, Anthropic, or any OpenAI-compatible endpoint — including self-hosted models, local inference servers, and third-party APIs like Groq, Together, or Mistral — and switch between local and cloud models in the app. This lets you use local models for privacy-sensitive tasks and cloud models when you want additional capability or speed.

Yes. ShadowAI is actively evolving, with ongoing work focused on improving generation quality, reducing latency, and expanding capabilities across text, audio, image, and video workflows. Future updates will also introduce more advanced orchestration and automation features.

Yes. Improving usability and customization is a priority. We're working on a clearer settings experience, a more intuitive interface, and more control over how the app looks and behaves. The current interface is functional, but it is not the final direction.

No. ShadowAI does not include telemetry, analytics, or user tracking. The app periodically contacts our update server to check whether a new version is available. No personal data or usage information is sent — only your current app version. Automatic update checks can be turned off in Settings. If you have Discord open, ShadowAI will connect to Discord's local Rich Presence IPC to display your activity status (e.g. "Using ShadowAI"). This connection is local — no data is sent to our servers. You can disable this in Settings if you prefer. Aside from the above, your data stays on your device unless you choose to connect to a cloud provider through your own API key. If you encounter issues, please contact us directly so we can investigate.

No. ShadowAI runs on CPU-only systems, though a GPU substantially improves performance for local inference, voice, and media generation. Supported backends: CUDA (NVIDIA), HIP/ROCm (AMD), Vulkan (cross-vendor), and DirectML (AMD/Intel integrated). The app detects your hardware automatically and picks the best one — you can override it in Settings. No GPU? You can still use the app fully through cloud providers.

ShadowAI can run multiple local AI components simultaneously — a local LLM loaded in memory via Ollama or llama.cpp, the audio runtime for voice and speech recognition, and the vision runtime for image or video generation. Each of these carries its own memory footprint, which adds up quickly. With 16 GB of RAM, you can run all features comfortably at the same time. With 8 GB, the app still works well for chat and agents, but running image or video generation alongside a loaded LLM may cause memory pressure. For media generation (image, video, music), GPU VRAM is just as important as system RAM. 6 GB of VRAM is a practical minimum for image generation; 8 GB or more is recommended for video.