Context Cooler MCP

Overview

Context Cooler is an MCP server that cuts agent token usage by 70-98% by running commands in sandboxed subprocesses, so only compact, intent-gated summaries ever enter the model's context window. Raw command output — build logs, search dumps, file reads — stays in the sandbox; the agent sees only what the intent filter determines it actually needs.

Beyond sandboxed execution, it supports 11 languages, FTS5 full-text search indexing for instant retrieval of previously seen output, batch execution for multi-command workflows, and 5 platform adapters for multi-messenger delivery. The result is an agentic toolchain that stays useful on long sessions without blowing the context budget.

The project is MIT licensed with zero callbacks, and is used by developers building with Claude Code, Cursor, and other agentic tools to keep token-hungry workflows sustainable.