An MCP server called Context Mode tackles the hidden problem of context window bloat in Claude Code by compressing tool outputs before they reach the model. Every MCP tool call dumps raw data—a Playwright snapshot costs 56 KB, 20 GitHub issues cost 59 KB—consuming up to 40% of your context within 30 minutes. Context Mode processes these outputs in sandboxes and returns only summaries, shrinking 315 KB of raw data to just 5.4 KB.

Built on the same principle as Cloudflare's Code Mode, which compressed tool definitions by 99.9%, Context Mode focuses on the other half of the problem: tool outputs. It supports 10 language runtimes, SQLite FTS5 with BM25 ranking for search, and batch execution. The tool includes smart snippets that extract relevant content around query terms rather than truncating arbitrarily, and progressive throttling that encourages batching queries after the first few calls. When installed as a plugin, it automatically routes large outputs through Context Mode without prompting.

Real-world benchmarks show dramatic savings: a Playwright snapshot drops from 56.2 KB to 299 bytes (99% saved), GitHub issues from 58.9 KB to 1.1 KB (98% saved). Session time before slowdown extends from ~30 minutes to ~3 hours, with context remaining at 99% after 45 minutes instead of 60%.