Show HN: AI Chat Terminal – Private data stays local, rest goes to cloud (github.com)

AI Chat Terminal is an open‑source Show HN project that gives you a hybrid chat: GPT‑4o in the cloud for general queries and a small local model (Qwen 2.5 Coder, ~5GB) for privacy‑sensitive operations. It routes user input through a two‑stage system: a <1ms keyword scan (action verbs like save/store/remember/show/delete in English/German/Spanish) and, if triggered, a local intent analysis by Qwen that decides whether to INSERT/SELECT/DELETE from an encrypted SQLite store. When Qwen confirms a local save the data never leaves your machine; SQLCipher (AES‑256‑CBC) protects the DB and macOS Keychain holds encryption keys. OpenAI is only contacted when no keyword is found or when Qwen classifies the message as a question. This approach matters because it gives explicit, user‑driven privacy control instead of unreliable automatic sensitive‑data detection: users choose when data is kept local, while still leveraging GPT‑4o for complex context and creativity. Technical implications: you need Ollama to run the local model, ~4.5–5GB disk for Qwen, Python and zsh on macOS 12+, and an OpenAI API key; chat history to OpenAI is ephemeral (deleted on exit or after 30 minutes) while keyword‑saved data persists locally until you delete it. The tradeoffs are clear: manual keyword activation reduces accidental cloud leaks but relies on users remembering to tag sensitive items; performance and NLP capability differ between the small local model and cloud GPT‑4o.