Help & support.
A local agent harness whose tokens never leave your Mac. Here's how to get help with it.
Email a human.
direct · no ticket queueEmail us about anything — a bug, a question about routing or model downloads, or a feature request. Because Coil runs entirely on your Mac, please don't include any private prompt content in your message; we can help without it.
Email support ↵The short answers.
privacy · router · memoryNo. Every prompt, every routing decision, and every generated token stays on your device. Coil runs open-weights language models locally through Apple MLX on Apple Silicon. There is no cloud LLM, no account, and no telemetry. The only network egress is a one-time, consent-gated download of the model weights from public mirrors after onboarding.
A small resident router model reads each prompt and dispatches it to a tier — a fast model for quick turns, a stronger model for hard work — with confidence-gated escalation between them. Every reply carries a receipt: which model handled it, what tier, what confidence, and how many tokens per second the run produced.
In a local memory store on your Mac, inside the app's sandbox container. Distilled memories are indexed with on-device embeddings so retrieval, pinning, pruning, and search all run locally. Nothing indexed leaves the machine, and one Settings action can wipe the store.
Open-weights language models supported by Apple MLX. Coil pins one fast router model and one strong worker by default, and the resident-model panel shows what is loaded, pinned, resident, or evicted against your Mac's unified-memory budget — so you can see the headroom before you route a large prompt.
An Apple Silicon Mac (M1 or newer) running macOS 14 (Sonoma) or later. On first launch, Coil downloads the router and default worker weights over HTTPS with your consent; after that, it works fully offline.
Coil is a one-time purchase of $14.99 on the Mac App Store. No subscription, no per-token metering, no in-app purchases.