Repolode
投稿する
mksglu

mksglu/context-mode

MCP server for AI coding agents that keeps raw tool output out of the context window and persists session state across compaction.

GitHub で見る公式サイト
スター
30,905
フォーク
2,302
未解決の Issue
356
最終コミット
2 時間前

これは何？

Context Mode addresses the context-window problem that AI coding agents hit when MCP tools return large raw outputs such as browser snapshots, issue lists or log files. It runs tool work inside a sandbox so only the final result enters the conversation, and the project reports a 98% reduction on a 315 KB example.

It also tracks session events such as file edits, git operations, tasks, errors and user decisions in SQLite. Events are indexed with FTS5 and retrieved with BM25 search, so after the conversation compacts the agent can resume with the relevant state. Sessions that are not continued are cleared at the next start.

The project is delivered as an MCP server and is wired into many clients through plugins, hooks and config files, including Claude Code, Gemini CLI, VS Code Copilot, Cursor, OpenCode, Codex CLI and OpenClaw.

主な機能

  • 01Runs tool output in a sandbox so raw data stays out of the context window.
  • 02Stores session events in SQLite and restores relevant ones after compaction.
  • 03Searches indexed content with FTS5 and BM25 ranking.
  • 04Provides ctx_execute and related tools for running scripts that return only results.
  • 05Uses platform hooks to route tool calls across supported AI coding clients.

ユースケース

  • Long coding sessions where MCP tools such as browser or issue fetchers flood the context.
  • Analyzing many source files by having the agent write a script that outputs only the summary.
  • Resuming work after the agent compacts its conversation without re-reading prior state.

トピック

#mcp#ai-agent#context-window#claude-code#coding-assistant#sqlite#llm-tools#antigravity#claude#claude-code-hooks

よくある質問

context-mode とは？
MCP server for AI coding agents that keeps raw tool output out of the context window and persists session state across compaction. Context Mode addresses the context-window problem that AI coding agents hit when MCP tools return large raw outputs such as browser snapshots, issue lists or log files. It runs tool work inside a sandbox so only the final result enters the conversation, and the project reports a 98% reduction on a 315 KB example.
context-mode は無料で使えますか？ライセンスは？
context-mode は Other ライセンスのオープンソースです。商用製品で使う前にライセンス条件を確認してください。
context-mode は今もメンテナンスされていますか？
活発にメンテナンス中。context-mode の最終コミットは 2026年10月19日 です。
context-mode は何に使えますか？
Long coding sessions where MCP tools such as browser or issue fetchers flood the context. Analyzing many source files by having the agent write a script that outputs only the summary. Resuming work after the agent compacts its conversation without re-reading prior state.
context-mode はどこで入手できますか？
ソースコードとインストール手順は GitHub にあります：https://github.com/mksglu/context-mode

概要はプロジェクトの README をもとに AI が生成しています。GitHub の数値は毎日更新されます。

関連プロジェクト