The problem
Most assistants make you choose: the convenience of a hosted chat, or the privacy of keeping your prompts on your own hardware. miii refuses the trade. It runs entirely on your machine and still does the things you actually want, tool use, document grounding, live search.
The decisions that mattered
- Two front ends, shared memory. A Next.js web UI and an Ink terminal app talk to the same local brain, so a conversation started in one continues in the other.
- LangGraph for agentic tool routing, so the assistant decides which skill to invoke instead of me hard-coding branches.
- Chroma-based RAG for document-aware answers, kept local.
- Pluggable skills defined as JSON, so extending the assistant does not mean touching the core.
- Tavily web search as an explicit opt-in, not an always-on data leak.
What it shows
miii is the product-shaped sibling of miii-cli: the same local-first conviction, applied to a general assistant instead of a coding agent. It is where I work out what a privacy-respecting AI product feels like end to end, from streaming NDJSON to slash commands to model switching.