stuntd: a local proxy that learns your LLM decisions and serves them at 22ms without an API key
Inevitable-Log5414 · reddit · 2026-09-23
A Reddit user open-sourced stuntd (Apache-2.0), a local proxy for closed-choice LLM decisions (choice, yes/no, score). It records your traffic, trains a small head on frozen-encoder Laya, runs in shadow mode, and only goes live once it matches your provider at the target agreement rate.
Why it matters
- Jev's hosted API costs 380ms per round trip; Laya answers in 22ms on a laptop GPU but scores 38% zero-shot on banking77 vs Jev's 76%
- After training: Snake plays 99.6% of moves locally, support triage answers 90% of questions locally, a coding-agent command gate goes from 45% to 97% accuracy (96% local)
Limits
- Closed decisions only; frozen encoder can't do arithmetic — "amount > X" scores 0.42 while a natural-language version hits 0.94
- Ships a 20-line PreToolUse hook that gates Claude Code shell commands with a local commandgate question
Related event: Open-Source Stuntd Learns Your LLM Decisions Locally(2 posts)→
More from coding & agent
- Rivet launches SQLite Forking: 20ms agent-thread forks with zero data copying — samgoodwin89 · 2026-11-11
- YouArt Desktop launches: AI agent edits video in CapCut, DaVinci and Premiere from client feedback — mhdfaran · 2026-11-11
- Boubou: one-command local voice AI that serves your phone and robots over Wi-Fi — LuozhuZhang · 2026-11-11
- Python 3.15 makes AttributeError actionable: suggests .append for .push mix-ups — KhuyenTran16 · 2026-11-11
- sentence-transformers maintainer releases new embedding pooling training method, code & checkpoints — tomaarsen · 2026-11-11
- droid-mcp turns your Android phone into an MCP server with 160+ tools for Claude Code — zvone1122 · 2026-11-11