Context Mode - An MCP Server That Cuts AI Coding Agents' Context Window by 98%

Context Mode is a local server designed to prevent raw data coming in from MCP tool calls from piling up directly in the context window.

• It isolates raw data in a sandbox, claiming 315KB → 5.4KB, i.e. a 98% reduction.
• It states that it extends session duration from about 30 minutes → 3 hours.
• It stores file edits, git operations, tasks, errors, and user decisions in SQLite + FTS5 and retrieves them via BM25 search to restore session continuity.
• Instead of having the LLM read data directly, it writes a script and receives only the results, reducing output tokens by 65–75% while maintaining technical accuracy, according to the explanation.
• It supports 14 platforms including Claude Code, Codex CLI, Gemini CLI, and Cursor, with session continuity support varying depending on the PreToolUse, PostToolUse, SessionStart, and PreCompact hooks.
• It provides 11 MCP tools such as ctx_execute, ctx_batch_execute, ctx_execute_file, ctx_index, ctx_search, and ctx_fetch_and_index, and applies Claude Code's permissions deny/allow rules to the sandbox as well.
• It emphasizes fully local execution, with no telemetry, cloud sync, or usage tracking, and the SQLite DB is stored in the home directory.
• The license is Elastic License 2.0.