Developer Builds Personal AI Gateway with Local Routing on macOS

On November 2, 2026, a developer who describes himself as working in B2B GTM and strategic marketing posted on V2EX about a personal AI gateway he built on a MacBook Pro. The project started around mid-2026, after what he called an out-of-control calling environment: a dozen-plus model subscriptions and API Keys across DeepSeek, NVIDIA NIM, DashScope, Gemini and others, with clients and scripts connecting directly and endpoints hard-coded into the codebase, making monthly spend impossible to account for. He framed the issue as a lack of control over invocations rather than a model-quality problem, and set out three goals: a single entry point so every client needs to be configured only once, intelligent routing that dispatches by task difficulty, and full auditability with a log of every call. The gateway now intercepts all local LLM calls, covering Cursor, Codex CLI, Claude Code and his own Agent content pipeline. A local 4B small model handles semantic routing: easy tasks go to cheap channels, while complex requests are forwarded to top-tier cloud models, and failed channels can be swapped out without users noticing. On the architecture side, the decision gateway listens on port 8081 as a self-written FastAPI service, kept running as a launchd daemon process, and performs five tasks before a request ever reaches a model, including outbound safety checks and local semantic routing. So far this setup is documented only in that single post, with no further developments reported.