mksglu/context-mode

Context Mode is an MCP server, a small program that plugs into AI coding agents like Claude Code, Cursor, and a list of others, and changes how those agents handle large pieces of data. The README explains the core problem: every time an agent calls a tool, the raw response goes straight into the model's context window. A web page snapshot taken with Playwright might be 56 KB, twenty GitHub issues might be 59 KB, a single access log file might be 45 KB. After half an hour of this, the README says, 40% of the working memory has been spent on data the model is no longer using, and when the conversation has to compact itself, the agent loses track of what it was doing. The tool addresses this in four ways according to the README. First, sandbox tools run external commands and keep the raw output out of context, replacing it with a small summary; one stated figure is 315 KB becoming 5.4 KB, a 98% reduction. Second, session continuity is tracked in a SQLite database with full-text search; when the conversation is compacted, only the events the model needs are retrieved via BM25 ranking rather than dumping everything back in. Third, what the README calls Think in Code asks the agent to write a script that returns only the answer rather than reading every file. Fourth, the project deliberately does not impose a writing style on the model's replies, citing benchmarks where strict brevity rules hurt coding performance. Installation is grouped by platform. The README documents 15 of them. Claude Code is the easiest case: two slash commands add the plugin marketplace and install context-mode, after which a SessionStart hook injects the routing instructions automatically and a doctor command validates the setup. Platforms with hook support get fully automatic routing enforcement; the others need a one-time routing file copied into the project. The MCP server ships eleven tools. Six are sandbox tools for running code and indexing data: ctx_batch_execute, ctx_execute, ctx_execute_file, ctx_index, ctx_search, and ctx_fetch_and_index. Five meta-tools cover statistics, diagnostics, upgrades, purging old data, and insights. Slash commands surface the same functions inside the agent. The README opens with logos of companies whose teams the project claims as users and shows a 570-point Hacker News thread as social proof. The license is the Elastic License v2.