Rlhf Feedback Loop

Captures thumbs up/down signals from AI agent interactions and turns them into permanent prevention rules that block repeated mistakes before they cost tokens. Built around the PreToolUse hook pattern, it intercepts risky tool calls like force pushes or destructive commands before they execute. The MCP interface exposes feedback capture, lesson promotion, and DPO export for fine-tuning workflows. Works with Claude Code, Cursor, Codex, and other MCP-compatible agents. Reach for this when you're tired of paying frontier model rates to fix the same hallucinated function call or dangerous command three sessions in a row. The underlying product is ThumbGate, which also ships a generated BRAIN.md artifact that consolidates approved lessons, guardrails, and gates into a single agent-readable context file.