# IgorGanapolsky/ThumbGate [Health: Active]

**Category:** 🤖 Coding Agents  
**Repository:** https://github.com/IgorGanapolsky/ThumbGate  
**GitHub Stars:** 29  
**npm Downloads (last month):** 2506  
**Views:** 2  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/igorganapolsky-thumbgate

## Description
MCP server that blocks AI coding agents from repeating mistakes — turns thumbs-up/down feedback into enforced pre-action gates via PreToolUse hooks. Install: npx thumbgate.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "thumbgate": {
    "command": "npx",
    "args": ["-y","thumbgate"]
  }
}
```

## Documentation

## What IgorGanapolsky/ThumbGate does

IgorGanapolsky/ThumbGate places a local decision point between an AI coding agent's proposed action and the tool that would execute it. It is designed to catch mistakes before they affect a checkout, expose a secret, or alter a repository.

By default, the gate blocks detected secret leaks and two classes of commands intended to disable ThumbGate itself: terminating the gate process and enabling its bypass environment override. Other high-risk patterns, including `rm -rf`, force-pushes, fetch-and-run commands, and direct edits to guardrails, produce a warning and are logged. Everything else is allowed unless a learned rule matches it.

Thumbs-down feedback can be recorded as a local lesson with context, the failed behavior, and the desired change. Repeated concrete failures can become prevention rules. These rules can promote a warning into a blocking gate without retraining the model.

## How it works

The enforcement path runs locally and does not require a cloud service. Agent integrations use either a native pre-tool hook or the MCP interface, depending on the client. Claude Code, Codex, and Gemini CLI use hard PreToolUse-style enforcement; Cursor, Cline, OpenCode, and generic MCP agents use the `gate_check` MCP tool as an advisory path.

Strict enforcement is controlled with `THUMBGATE_STRICT_ENFORCEMENT=1`. In that mode, checks that would normally warn and log become hard denials. For example, a saved lesson about avoiding `DROP` on production tables can cause a later matching proposal to be rejected before execution.

## Setup and configuration

Install the command-line package through npm and initialize the agent-specific wiring:

```bash
npx thumbgate init
npx thumbgate doctor
```

`doctor` verifies that the PreToolUse hook is installed. To use ThumbGate as an MCP server, start the stdio transport instead:

```bash
npx -y thumbgate serve
```

`npx -y thumbgate mcp` is an equivalent MCP command. Do not substitute `npm start` for MCP setup; that command starts the hosted HTTP API rather than the agent-facing stdio server.

A local dashboard can be opened with `npx thumbgate dashboard --open`. Capture feedback with a command such as:

```bash
npx thumbgate capture --feedback=down --context="Never run DROP on production tables" --what-went-wrong="agent proposed DROP" --what-to-change="require review for DROP"
```

## Tools and capabilities

- `init` configures supported agent hooks or MCP wiring.
- `doctor` checks whether the PreToolUse hook is connected.
- `serve` and `mcp` expose the stdio MCP server.
- `gate_check` evaluates a proposed action through MCP.
- `capture` records accepted feedback as a local lesson.
- `eval` converts accepted feedback into reusable evaluation cases and local proof reports.
- `dashboard` opens local enforcement statistics.
- Strict mode changes warning verdicts into denials.

## Limitations and notes

ThumbGate intercepts tool calls at runtime; it does not update model weights. The native hook path provides hard enforcement for Claude Code, Codex, and Gemini CLI. Cursor, Cline, OpenCode, and other MCP agents use the advisory MCP `gate_check` route rather than the same native hard-hook path.

The default policy is intentionally mixed: only certain secret-leak and self-disable classes are blocked immediately, while several destructive or risky command classes warn and log until strict enforcement is enabled. MCP registries and clients must launch the stdio server, not the HTTP API.

_Full upstream README: https://allmcps.com/mcp/igorganapolsky-thumbgate/readme_