Introducing Middleman — The Precision Context Refiner for MCP Pipelines

[JithunMethusahan]

Pre-submission Checklist

• This post follows the MCP community guidelines

What would you like to share?

Hey everyone 👋

I'm excited to share a tool I built to solve a problem that's been quietly draining tokens from every MCP-powered workflow: context bloat.

The Problem

When your AI agent fetches a Wikipedia article or scrapes a webpage via MCP, it receives the entire raw payload — often 15,000–25,000 tokens of noise for maybe 800 tokens of useful signal. You pay for all of it.

What Middleman Does

Middleman is an MCP server that sits between your data source and your LLM. It intercepts the raw data, refines it using Gemini 1.5 Flash, and delivers a clean, XML-structured summary.

Real numbers:

Source	Raw Tokens	After Middleman	Reduction
Wikipedia article	~25,000	~800	96.8%
Reddit thread	~15,000	~600	96%
System logs (5MB)	~500,000	~1,200	99.7%

Key Features

• Fetch & refine any URL
• Process local .txt, .log, .md files
• Surgical focus queries ("only extract technical specs")
• Strict XML output schema for machine-to-machine reliability
• Built on FastMCP, MIT licensed

Get Started

git clone https://github.com/JithunMethusahan/middleman.git
cd middleman && pip install -r requirements.txt

Repo: https://github.com/JithunMethusahan/middleman

I'd love feedback from the community — especially if you test it with your own MCP stack. Happy to answer questions below!