mksglu/context-mode
Sandboxes MCP tool output so AI coding agents stop drowning their context window
An MCP server that keeps raw tool output — a Playwright snapshot, a batch of GitHub issues, a log file — out of the context window instead of dumping it in; its own example goes from 315 KB to 5.4 KB, a 98% cut. Tracks file edits, git operations, tasks and errors in SQLite, and after a conversation compacts it retrieves only what's relevant via BM25 search instead of replaying the whole history. Pushes a "think in code" pattern — the agent writes a script to compute an answer instead of reading dozens of files into context — across 17 supported clients including Claude Code, Cursor and OpenClaw. It explicitly doesn't touch how the model phrases its final answer — brevity and formatting stay up to your own prompts.
What holds up
- +Cuts raw tool output by 98% in its own example — 315 KB down to 5.4 KB
- +Session data is deleted immediately if you don't --continue — no memory hanging around
- +Works across 17 clients including Claude Code, Cursor and OpenClaw via MCP + hooks
Mind the limits
- −Only automatic on hook-capable platforms — others need a one-time manual routing file copy
- −Adds SQLite session tracking and a mandatory scripting workflow, extra moving parts to trust
Featured here? Take the badge
Put it on your site — it links back to this review. Free for every listed product, always.
<a href="https://stillworks.dev/products/p/mksglu-context-mode/"><img src="https://stillworks.dev/badge/mksglu-context-mode.svg" alt="Picked by StillWorks" width="250" height="54"></a>