What it does Context Mode is an MCP server that intercepts the flood of raw data from tool calls — Playwright snapshots, GitHub issues, log files — and keeps it out of the LLM’s context window. It replaces verbose output with sandboxed execution results, tracks session state in a local SQLite database with FTS5 search, and enforces a “think in code” pattern where the agent writes scripts instead of reading files into context. It supports 15+ platforms including Claude Code, Gemini CLI, Cursor, and Zed.

The interesting bit The project treats context compression as a data-routing problem, not a prompting problem. When the conversation compacts, Context Mode doesn’t dump history back into the window — it retrieves only relevant events via BM25 search. The “think in code” mandate is unusually strict: the LLM must generate scripts that console.log() results, turning 47 file reads into one 3.6 KB execution. The README explicitly rejects prose-style enforcement, citing evidence that aggressive brevity prompts degrade reasoning benchmarks.

Key highlights

• Claims 98% context reduction (315 KB → 5.4 KB) via sandboxed tool output
• Session continuity through SQLite + FTS5 indexing with BM25 retrieval
• 11 MCP tools: six sandbox executors (ctx_execute, ctx_batch_execute, etc.) plus five meta-tools (ctx_stats, ctx_doctor, ctx_insight)
• Hook-based automatic routing on Claude Code and Gemini CLI; manual config file for other platforms
• Optional status line showing real-time savings per session
• ctx_insight opens a local analytics dashboard with 90 metrics across 23 event categories

Caveats

• License is ELv2 (Elastic License v2), not open source
• Platform coverage varies: Claude Code gets full plugin marketplace install, others require manual config file edits or hook setup
• The 98% reduction figure is self-reported from the README without independent verification
• Session data is deleted on fresh starts unless --continue is used, which could surprise users expecting persistent history

Verdict Worth trying if you hit context limits daily with agentic coding tools — especially Claude Code users who can install via marketplace. Skip if you need true open source or if your workflow doesn’t involve heavy tool-call churn.