What is context-kernel?

Context kernel is a tool built for AI coding assistants like Claude Code, aimed at solving a common problem, these agents often waste a large part of their limited context window on information that has nothing to do with the task at hand. Instead of giving the assistant an entire repository or every tool output, this project narrows things down to just the files and information that actually matter for the specific task being solved. The core idea is that what counts as relevant context always depends on the task itself. There is no single fixed set of files that matters, it changes based on what you are trying to fix or build. Context kernel takes a repository and the tool outputs an agent has already produced, and reduces them down to a smaller, focused set that still contains everything needed to answer correctly. The project describes this as different from simple compression, since the output stays in its original, readable form rather than becoming shorter but harder to understand. The project reports specific measured results rather than estimates. On one full real world coding session it reduced total token usage by 79 percent from start to finish. On two separate test sets built from real bugs in the pandas and Django codebases, it kept the piece of code needed to actually fix the bug inside the reduced working set 100 percent of the time across 120 combined test cases. It also ships with several hundred automated tests covering this behavior across Linux, Windows, and macOS. Context kernel works as a native plugin for Claude Code, with support for Codex and a general purpose interface called MCP that other AI coding assistants can call into as well. It runs using only Python's standard library, meaning it needs no extra installed dependencies and makes no calls to any external API or paid service, so everything it does happens locally within your own coding session.