// the find
johnbean393/Sidekick
A native macOS app that allows users to chat with a local LLM that can respond with information from files, folders and websites on your Mac without installing any other software. Powered by llama.cpp.
Sidekick is a native macOS app for running local LLMs via llama.cpp, with RAG over your own files, agentic tool-calling, and a web search integration. It targets Mac users who want a ChatGPT-like experience without sending data to the cloud — researchers, students, and developers who work with personal documents. Apple Silicon only.
Ships its own llama.cpp binaries so there's genuinely zero setup beyond downloading the app — no Homebrew, no Python env, no Ollama. The 'Experts' abstraction (named collections of files/folders/URLs) is a clean way to scope RAG context without fumbling with per-message attachments. Speculative decoding support and Metal acceleration are wired up properly, so on M-series chips it's actually fast. The agentic loop (sequential tool calls until done) is implemented end-to-end, not just a demo — the Contacts/Calendar/email draft integration uses real system APIs.
GraphRAG is present in the tree but there's no documentation on how it interacts with the existing embedding-based retrieval — likely incomplete or experimental, which matters if you're choosing this for anything serious. The RAG pipeline uses a local similarity index (SimilarityIndex extension) rather than a proper vector store, so performance with large document collections is unknown and probably degrades. Image generation requires Apple Intelligence (macOS 15.2+) and falls through silently on older systems — the auto-detection UX sounds slicker than it probably is in practice. The project is a one-person effort with 143 forks and no visible test suite, so the agentic tool-calling reliability is entirely dependent on what the bundled llama.cpp version can handle with each model.