thumbgate

# ThumbGate Pre-action gates that stop AI coding agents from repeating known mistakes. ## Quick Start ```bash npx thumbgate init ``` This installs the MCP server and wires it into your agent's tool configuration. No API keys required for the free tier. Or install globally: ```bash npm install -g thumbgate thumbgate init ``` ### MCP Configuration Add to your agent's MCP config (e.g., `claude_desktop_config.json` or `.cursor/mcp.json`): ```json { "mcpServers": { "thumbgate": { "command": "npx", "args": ["-y", "thumbgate"] } } } ``` ## How It Works ### Feedback Capture When an agent action succeeds or fails, capture structured feedback: - **Thumbs up**: Records what worked, tags it, and stores it as a reusable pattern. - **Thumbs down**: Records the failure context, what went wrong, and what to change. Repeated failures auto-promote into prevention rules. ### Prevention Rules After a failure pattern repeats (default: 2 occurrences), ThumbGate auto-generates a prevention rule. These rules are injected into agent context before every tool call, blocking the known-bad pattern before it executes. ### Pre-Action Gates Gates intercept tool calls via the MCP PreToolUse hook. Each gate checks the proposed action against: 1. Prevention rules generated from past failures 2. Thompson Sampling confidence scores (adaptive sensitivity) 3. LanceDB vector similarity to known-bad patterns If a match is found, the gate blocks execution and surfaces the p...