context-mode

Context Mode is an MCP server and set of hooks that keeps raw tool output out of an AI coding agent's context window. Its starting observation is that every tool call dumps data into context: a Playwright snapshot, a batch of GitHub issues or an access log can each cost tens of kilobytes, and after half an hour a large share of the window is gone. Context Mode runs such tools in a sandbox and returns only what the agent needs.

It tackles four parts of the problem. Sandboxed tools cut raw output dramatically (the project reports around 98% on its examples). Session continuity tracks file edits, git operations, tasks, errors and decisions in SQLite and indexes them, so work survives context compaction without being dumped back into the window. "Think in code" encourages the agent to write a small script that computes an answer instead of reading dozens of files into context. And it deliberately does not dictate how the model writes its final answer. It works across many coding platforms through MCP and hooks, with the deepest integration as a Claude Code plugin.

Getting started: in Claude Code, add the `mksglu/context-mode` plugin marketplace, install the plugin and run `/context-mode:ctx-doctor` to verify. For other platforms, install the global binary with `npm install -g context-mode` and follow the platform-specific setup.