🔥 Hot Repo: #1 on HN — Slashes Agent Context Windows by 98%

One-liner — context-mode is an MCP server that sandboxes tool output before it reaches the context window, cutting token usage by up to 98% across 15 AI coding platforms. - Repo: mksglu/context-modehttps://github.com/mksglu/context-mode - Stars: ⭐ 14,783 - Language: TypeScript - License: Elastic License v2 ELv2 --- What It Does Every MCP tool call dumps raw data into your context window. A Playwright DOM snapshot costs 56 KB. Twenty GitHub issues cost 59 KB. One access log — 45 KB. After 30 minutes of active agent work, 40% of your context window is consumed by tool output alone. context-mode intercepts those calls via PreToolUse/PostToolUse hooks, routes the raw output to a local SQLite store, and returns only a compact summary — shrinking 315 KB down to 5.4 KB. The agent still gets the answer; the context window doesn't pay for it. Why It's Blowing Up context-mode hit 1 on Hacker News with 570+ points shortly after launch, driven by a pain point that no existing tool was attacking: the output side of the context problem. LLM context managers like CLAUDE.md rules and compaction handle what goes in to the model; nothing was handling what tool calls pour back into the window after execution. The cross-platform story adds fuel. Most context tools are Claude Code-specific. context-mode ships native hook configs for Gemini CLI, VS Code Copilot, Codex, Cursor, Kiro, OpenCode, Zed, and eight others — one MCP server, one interface, every major agent runtime. v1.0.135 shipped today May 15, 2026 with critical fixes including a Pi adapter regression that caused all ctx calls to fail after MCP idle timeout, plus a Linux SIGSEGV escape hatch for Bun/Node compatibility edge cases. At 14,783 stars in under three months since its February 2026 debut, the growth rate is steep. The repo claims adoption at teams inside Microsoft, Google, Meta, Amazon, NVIDIA, and ByteDance — though these are self-reported badges. Key Features - Sandbox tools — Six MCP tools ctxexecute,