ntk

What problem does it solve?

NTK solves the problem of excessive LLM context and cost by routing each request to the minimum-depth multi-agent pipeline needed for the task’s complexity, while compressing and selectively sharing information across agents.

Core Features & Use Cases

• Adaptive complexity routing: Automatically selects the execution depth (direct → light → standard → full) to avoid over-processing simple tasks.
• Need-to-know information sharing: Reduces token waste by only providing each agent the minimal context it needs, rather than broadcasting full history.
• Information compression with backtracking (Tee): Compresses gathered information and can recover original text if verification fails, improving reliability without always paying full-context costs.
• MCP-first integration: Exposes capabilities as an MCP server so you can call NTK from MCP-compatible AI clients (VS Code Copilot, Claude Desktop, OpenClaw, etc.).
• Dual-model cost isolation: Uses a strong model only for the highest-depth planning stage, while low-depth execution relies on a cheaper model tier.

Use case: When you ask for a simple function or translation, NTK stays in direct mode; when you ask for architecture design, it escalates to full mode with planning, verification, and compressed research steps.

Quick Start

Install NTK, create/edit a .env file with your OpenAI-compatible API endpoint credentials, then start the MCP server and call ntk_run from your MCP client.