Why this matters
Large LLM-driven coding sessions quickly exhaust context windows when tools (Playwright, CLI dumps, logs, web fetches) stream raw output into the conversation. That forces frequent compaction and loses in-progress edits, tasks, and decisions. Context optimization therefore isn't just an efficiency problem — it's a continuity and correctness problem for agent workflows. Context Mode solves this by moving heavy processing out of the context window and into a local, sandboxed MCP server that indexes results for on-demand retrieval. (github.com)
What Sets It Apart
-
Sandbox-first tooling: tool calls run in isolated subprocesses; only stdout (or intent-filtered snippets) re-enters the conversation, preventing large dumps like Playwright snapshots or access logs from overflowing the model context. This yields benchmarked savings such as a ~99% reduction on snapshots and ~98% on multi-issue dumps. (github.com)
-
Indexed, relevance-driven retrieval: raw outputs are chunked and stored in a local SQLite FTS5 index with BM25 ranking, Porter stemming, trigram fallback, proximity reranking, and fuzzy correction — the model only receives concise, highly relevant snippets via BM25/RRF merging rather than raw files. (github.com)
-
Session continuity across compaction: platform hooks capture file edits, git ops, errors, and user decisions into the DB and build compact session snapshots so the agent can resume without re-prompting after context compaction. The project provides adapters/configs for major agent platforms and a small set of sandbox MCP tools (execute, execute_file, index, search, fetch_and_index, batch_execute). (github.com)
Who it's for — and trade-offs
Great fit if you run LLM-driven code workflows or integrate AI coding agents into developer tools and want to preserve session state, reduce token usage, and keep sensitive outputs off the context stream. It benefits teams using Claude Code, Gemini CLI, VS Code Copilot, Cursor, OpenCode, and similar platforms (adapters and hooks available or planned). (github.com)
Look elsewhere if you need a hosted/cloud service (Context Mode is local, ELv2-licensed and privacy-first) or prefer a pure UI application — this is protocol-level infrastructure intended for embedding at the MCP/hook layer, not a front-end chat product. Also expect some setup when adding adapters for niche platforms. (github.com)
