Tool Guardian - Reducing Token Overhead for MCP Tool Definitions and Results

In MCP tool calling scenarios with local models (such as Ollama, LM Studio, vLLM, or llama.cpp), introducing multiple MCP servers can cause tool definitions to quickly fill up the context window. Tool Guardian, as an MCP routing layer, mitigates this issue by reducing explicit tool definitions.

What It Is

Tool Guardian is an MCP routing tool that sits in front of your MCP servers, replacing dozens of specific tool definitions with 6 generic tools (8 when used with Skills). It uses an “on-demand discovery” mechanism to prevent tool definitions from filling the window before the model reads content.

This plugin is maintained by LuminariSoftwares and provides a DeepSeek Harness Native Bundle, aimed at solving context overflow issues caused by tool calls.

Core Features

  • MCP Routing: Provides a unified entry point for managing multiple MCP servers.
  • DeepSeek Harness Native Bundle: Direct integration as a DSH plugin.
  • Tool Schema Optimization: Reduces initial context usage through progressive disclosure.
  • Tool Result Archiving: Archives tool results to prevent them from accumulating in the conversation window.
  • Tool Group Cost Calculation: Calculates the cost of tool calls.
  • Progressive Disclosure: Shows specific tools only when the model needs them.

Installation

Use the official installation command to add the Bundle to your DSH configuration:

dsh plugin --profile web add dsh-tool-guardian

Notes and Compatibility

Before installing, make sure your environment meets the following dependency requirements:
* DeepSeek Harness (DSH): 0.1.2-alpha.2 or later
* Node.js: ^22.19 or >=24
* Python: 3.9+

This plugin is compatible with the following client environments:
* Claude Code
* OpenCode
* Cline
* Continue

Value Summary

Tool Guardian compresses tool definitions that originally take about 20,005 tokens down to 1,209 tokens, and supports on-demand archiving and invocation of tool results. Used together with Context Guardian, it helps solve the problem of the context window being filled by tool definitions and results.

羽毛球分组比赛记分
小程序二维码

欢迎使用《羽毛球分组比赛记分》微信小程序

Xiaoye