Your Mac is the cheapest AI provider you will ever use.

Rumrunner is a background daemon for macOS. It serves an OpenAI-compatible endpoint on localhost, so the tools you already use — Cline, Continue, aider, Zed — point at it by changing one setting. Every request is routed to whichever engine can actually do that job, and you get an itemised receipt of what it saved.

Three lanes, one endpoint

Ashore — a model running on this Mac, via MLX or llama.cpp. Free.

Nearshore — cheap hosted open-weight models. Fractions of a cent.

Offshore — Anthropic, OpenAI, DeepSeek, on your own key. Real money, used only when it earns it.

The receipt is the product

Every routing decision is logged with its inputs, on your machine. The ledger adds them up and tells you what you actually saved against paying frontier prices for everything — including the months when the answer is “not much”.

Example — last 30 days

2,979 requests

Ashore on this Mac 1,84762%

Nearshore hosted open-weight 80427%

Offshore your own key 32811%

All 2,979 sent offshore, at list$472.10

What you actually paid$53.50

Saved$418.60

Illustrative figures. Your numbers will be your own.

What it will not do

Your prompts stay on the machine unless a request explicitly routes to a hosted lane. Telemetry is aggregate counts. No prompt content is ever shipped anywhere.
Frontier tokens are billed at cost, with zero markup. You bring your own key and pay Anthropic or OpenAI directly. We never take a cut of your inference.
API keys live in the macOS Keychain. Never in a config file, never in a log, never in an error message.
Privacy fences are absolute. A request whose context touches a fenced path runs locally, whatever the router would otherwise have scored it.