🧰 K3 MCP Toolbox

"Give your AI Agent the tools it actually needs." A production-ready Model Context Protocol (MCP) Server for Windows & Docker.

⚡ Quick Start (The "30-Second" Rule)

Installation

pip install -r requirements.txt

Run Server

# Exposes the server on stdio (works with Claude Desktop / Cursor)
python server.py

Configure (Claude Desktop / Cursor) Add this to your claude_desktop_config.json:

{
  "mcpServers": {
    "k3-toolbox": {
      "command": "python",
      "args": ["/absolute/path/to/k3-mcp-toolbox/server.py"]
    }
  }
}

🛠️ The Tools

1. sequential_thinking (The Logic)

A "System 2" thinking module. Allows agents to break complex problems into dynamic steps, revise previous thoughts, and branch constraints.

• Based on Protocol 310.

2. mrl_search (The Brain)

Matryoshka Representation Learning (MRL) indexer.

• Zero-Setup: Uses local storage (no vector DB required).
• SOTA: Configured for text-embedding-004 (Gemini) by default.
• Fast: Funnel search implementation (64-dim shortlist -> 768-dim rerank).

3. kill_zombies (The Ops)

Windows Only. Instantly terminates stuck Chrome/Chromedriver processes and releases file locks. Essential for Selenium/Playwright agents.

🤖 Agent Context (AI-Ready)

This repository includes a context.md file optimized for LLM consumption. If you are using Cursor or Windsurf, simply @context.md to give your AI full understanding of this codebase.

🤝 Contributing

We follow the K3-9000 Growth Protocol.

• Hygiene: No .env commits.
• Style: Type hints required.
• Tests: Must pass test_mcp_health.py.

License: MIT