If your app asks an LLM the same routing, triage, or moderation question thousands of times a day, stuntd is the local proxy that learns those answers and stops paying for them. It sits in front of OpenAI or Anthropic as a drop-in replacement for the provider base URL, recording typed requests while relaying everything else byte for byte, and distils each decision site into a small head on a frozen Laya encoder. Once trained, the head answers in about 20 milliseconds with no tokens and no round trip; when its calibrated confidence falls below the site threshold, or the novelty gate flags the input, the question goes back to the provider instead of being guessed at. It speaks four protocols on one daemon: Jev System One with choice, score, and noul questions, the OpenAI Decisions API, Chat Completions, and the Anthropic Messages API, and every response carries an X-Stuntd header showing who answered. Run it keyless and the base Laya checkpoint answers zero-shot locally, put the paid Jev API in front to teach from real traffic, or import labelled rows and train directly. A few thousand rows close the gap: in the banking demo, the risk question climbs from 30.0 to 72.5 percent. A shipped Claude Code PreToolUse hook answers "may this command run" in about 50 milliseconds, without sending every command to a provider. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. Apache-2.0 licensed.