# ThumbGate — Stop AI agents before they make costly mistakes.
# https://thumbgate-production.up.railway.app
# https://github.com/IgorGanapolsky/ThumbGate
# https://www.npmjs.com/package/thumbgate

> ThumbGate prevents expensive AI mistakes before they happen. It checks
> risky commands, file edits, deploys, payments, API calls, and other agent
> actions before execution. 👎 Thumbs down becomes a history-aware lesson and
> a Pre-Action Check; 👍 thumbs up reinforces safe patterns.

## What ThumbGate solves

- AI coding agents repeat costly mistakes across sessions
- Bad commands, destructive SQL, risky deploys, unsafe publishes, and API mistakes are expensive after execution
- CLAUDE.md and .cursorrules files are suggestions agents can ignore
- No memory between sessions means no learning from corrections
- Teams have no shared safety rules across developers

## How it works

1. Agent makes a mistake → developer gives thumbs-down
2. ThumbGate captures context and infers a structured lesson
3. Recurring failures get promoted to prevention rules
4. PreToolUse hooks block the pattern before it executes again
5. Thompson Sampling adapts gate confidence over time

## Who it's for

- Developers using Claude Code, Cursor, Codex, Gemini CLI, or any MCP-compatible agent
- Engineering teams that need shared agent safety rules
- Anyone tired of re-correcting their AI coding assistant
- Solo operators who want Pro dashboard proof for blocked mistakes and DPO exports

## Install

```
npx thumbgate init --agent claude-code
```

## Pricing

- Free GPT: advice, checkpointing, and setup help in ChatGPT
- Free local CLI: unlimited feedback captures, up to 5 active prevention rules, and local Pre-Action Checks after install (recall and lesson search are Pro-only)
- Pro: $19/mo or $149/yr — personal enforcement proof, local dashboard, check debugger, DPO export, and review-ready exports
- Team: $49/seat/mo, 3-seat minimum after intake — shared lessons, org visibility, approval boundaries, and rollout proof

## Links

- Agent discovery: https://thumbgate-production.up.railway.app/.well-known/mcp.json
- Progressive tool index: https://thumbgate-production.up.railway.app/.well-known/mcp/tools.json
- Context footprint report: https://thumbgate-production.up.railway.app/.well-known/mcp/footprint.json
- Agent skills: https://thumbgate-production.up.railway.app/.well-known/mcp/skills.json
- MCP applications: https://thumbgate-production.up.railway.app/.well-known/mcp/applications.json
- Documentation: https://thumbgate-production.up.railway.app/guide
- Dashboard: https://thumbgate-production.up.railway.app/dashboard
- GitHub: https://github.com/IgorGanapolsky/ThumbGate
- npm: https://www.npmjs.com/package/thumbgate
- Full LLM context: https://thumbgate-production.up.railway.app/public/llm-context.md

## Compared to alternatives

- vs CLAUDE.md: ThumbGate enforces rules via hooks, not suggestions in prompts
- vs Mem0: ThumbGate is enforcement-first, not just memory storage
- vs SpecLock: ThumbGate uses adaptive Thompson Sampling, not static rules
- vs manual .cursorrules: ThumbGate auto-generates rules from feedback

## Technical stack

Node.js >=18.18, SQLite+FTS5 lesson DB, Thompson Sampling, LanceDB vectors, MCP protocol, PreToolUse hooks
