Context Mode Solves Claude Code Context Window Exhaustion — 98% MCP Output Reduction

A growing challenge among developers using AI coding agents has been MCP (Model Context Protocol) tool outputs consuming context windows, causing sessions to slow down after about 30 minutes. Context Mode, developed by Mert Köseoğlu, solves this problem by compressing MCP tool outputs by up to 98%.

The Problem: Context Window Exhaustion from MCP Outputs

MCP has become the standard tool protocol for AI coding agents, but each tool call consumes context window tokens. The problem is particularly severe on the output side, where the following data is added to context for every tool call:

Data Type	Context Consumption
Playwright snapshots	56 KB
GitHub Issues (20 items)	59 KB
Access logs (1 item)	45 KB
CSV analysis (500 rows)	85 KB

After 30 minutes of a session, 40% of the context is lost, degrading session longevity and response quality.

Context Mode’s Solution

Context Mode acts as an MCP server that receives raw data from MCP tools and returns only the minimum information AI needs to context. Compression results:

Data Type	Before	After	Reduction
Playwright snapshots	56 KB	299 B	99.5%
GitHub Issues (20 items)	59 KB	1.1 KB	98%
Access logs (500 items)	45 KB	155 B	99.7%
CSV analysis (500 rows)	85 KB	222 B	99.7%

For entire sessions, 315 KB compresses to 5.4 KB, extending session runtime from approximately 30 minutes to about 3 hours before context degradation.

Technical Approach

Context Mode consists of two main components:

1. Sandbox

Each execute call runs in an isolated subprocess. Raw data (logs, API responses, snapshots) never enters context—only stdout is returned, preventing unnecessary data from bloating the context.

2. Knowledge Base

Chunks markdown by headings and provides BM25 ranking search via SQLite FTS5. Returns code blocks precisely, keeping context information minimal yet sufficient.

Installation

# MCP only
claude mcp add context-mode -- npx -y context-mode

Or install via Claude Code’s Plugin Marketplace. Supports 10 language runtimes (JS/TS/Python/Shell/Ruby/Go/Rust/PHP/Perl/R), with 3-5x speedup when using Bun.

Why It Was Needed

The idea came from Köseoğlu, who operates the MCP Directory & Hub. “Everyone builds tools, but nobody addresses the output side. We process 100,000 requests daily at MCP Hub. When Cloudflare announced Tool Mode for compressing tool definitions, we tackled the reverse—compressing tool outputs.”

Conclusion

Context Mode addresses the often-overlooked “output side efficiency” in an era where MCP standardization is advancing. Efficient context window utilization will become increasingly important for future AI coding agent development.

Repository: github.com/mksglu/claude-context-mode (MIT License)