context-mode cuts agent context by 98 percent
context-mode is an MCP server trending hard on GitHub today — 9.4K stars, 302 added in 24 hours, and the reason is in the numbers. A Playwright snapshot normally costs 56 KB of context. Under context-mode it becomes 299 bytes. Twenty GitHub issues go from 59 KB to a handful. Access logs drop from 45 KB to whatever you query for. That is the 98 percent reduction the README promises.
Three mechanisms are doing the work. One, tool outputs run in isolated subprocesses and only the answer hits context. You asked how many functions in the file? The agent writes a script that outputs the number, not the whole file. Two, every file edit, git operation, and decision gets written to a SQLite database with FTS5 indexing and BM25 ranking. When the agent compacts its context, it retrieves relevant indexed events instead of dumping raw history back in. Three, code-first analysis — replace ten tool calls with one script.
Works on Claude Code, Gemini CLI, VS Code Copilot, Cursor, plus eight more. 24-hour TTL caching for URLs so fetched pages do not burn context twice. The repo claims Microsoft, Google, Meta, Amazon, IBM, and NVIDIA teams already using it. The v1.0.89 release hit April 14 but the daily star growth shows adoption is still compounding.
The real insight is that context is the actual agent bottleneck, not reasoning. A model that was smart at minute one is dumb by minute thirty because 40 percent of its context is gone to tool-output sludge. Offloading to a subprocess plus indexed retrieval is the same pattern Unix pipes used forty years ago — keep the big stuff in files, pass small handles around. It took the LLM world a surprisingly long time to figure this out.
https://github.com/mksglu/context-mode
← Back to all articles
Three mechanisms are doing the work. One, tool outputs run in isolated subprocesses and only the answer hits context. You asked how many functions in the file? The agent writes a script that outputs the number, not the whole file. Two, every file edit, git operation, and decision gets written to a SQLite database with FTS5 indexing and BM25 ranking. When the agent compacts its context, it retrieves relevant indexed events instead of dumping raw history back in. Three, code-first analysis — replace ten tool calls with one script.
Works on Claude Code, Gemini CLI, VS Code Copilot, Cursor, plus eight more. 24-hour TTL caching for URLs so fetched pages do not burn context twice. The repo claims Microsoft, Google, Meta, Amazon, IBM, and NVIDIA teams already using it. The v1.0.89 release hit April 14 but the daily star growth shows adoption is still compounding.
The real insight is that context is the actual agent bottleneck, not reasoning. A model that was smart at minute one is dumb by minute thirty because 40 percent of its context is gone to tool-output sludge. Offloading to a subprocess plus indexed retrieval is the same pattern Unix pipes used forty years ago — keep the big stuff in files, pass small handles around. It took the LLM world a surprisingly long time to figure this out.
https://github.com/mksglu/context-mode
Comments