Every token is generated on your Mac.
Coil runs open-weights language models through Apple MLX on Apple Silicon. Model weights download once, with your consent, from public mirrors. After that, Coil works offline for the lifetime of the app. No cloud LLM, ever.
The right model for each turn.
A small resident router reads every prompt and dispatches it: quick questions to a fast model, hard work to a strong one. Each answer carries its receipt: which model, what tier, what confidence, how fast.
It remembers on a disk you control.
Conversations distill into a local memory store that is searchable, pinnable, inspectable, and deletable. Retrieval, not reload: only what matters re-enters the context window, and you can watch the budget.
Per-turn tier dispatch between a pinned fast router model and a strong worker, with confidence-gated escalation. Every reply shows its routing receipt.
Live table of loaded models (pinned, resident, evicted) with resident memory and headroom against your Mac's unified-memory budget.
Distilled memories with on-device embeddings; search, pin, and prune from the app. Nothing indexed leaves the machine.
Index a codebase from the CLI into a local symbol graph. PageRank-ranked repo maps feed the agent structure, not source, and the app renders that graph, a human-gated knowledge wiki, and a local usage ledger.