context-mode Cuts MCP Tool Output From 315KB to 5.4KB

context-mode cut 315KB of MCP tool output to 5.4KB, a 98% drop, by the maintainer's own count. For teams burning through context windows on agent tool calls, that's fewer re-reads after every compaction. The repo cites a 56KB Playwright snapshot and 59KB from twenty GitHub issues as the kind of routine waste it's built to cut.

The project reached 19,893 stars and 1,430 forks on GitHub, licensed under Elastic License 2.0 and pushed on August 16.

The mechanism matters more than the percentage. Tool output gets sandboxed rather than dumped straight into context.

Every file edit, git operation, task, and error gets logged to SQLite. The project pulls that history back later through FTS5 and BM25 search, instead of re-reading the same files after a context compaction.

The project also enforces this routing across 17 agent platforms and trims filler the model tends to generate on its own.

The headline example: a task that used to take 47 Read() calls and 700KB now runs as one ctx_execute() script returning 3.6KB.

I don't know how that number holds up outside the repo's own test cases. GitHub doesn't show an independent benchmark, and the figure comes from the maintainer choosing which examples to publish.

The transferable idea

The part worth taking, even if you skip the install, is Think in Code. The model writes a script that computes the answer, instead of reading raw files and computing it itself. That's a pattern any agent setup can borrow, no matter what's managing the context window.

A related project on the same token-blowup theme, watermarks-remover, documented one case that dropped from 9,894 tokens to 12. It restructured how an agent skill reasoned about a problem. The instinct is the same: stop paying the model to read what it doesn't need.