跳到正文
DEV Community· Pumba·· 11 天前AI 評分45

CodeGraphy v2.1:本地上下文智能平台及 MCP Server,以互動式架構圖減少 85% LLM token 用量

I built CodeGraphy v2.1 - a Local Context Intelligence Platform & MCP Server with an interactive architecture graph that cuts LLM token usage by 85%

AI 導讀

CodeGraphy v2.1 是本地上下文智能平台及 MCP Server,透過互動式依賴圖和混合上下文處理,減少 85% LLM token 用量。它內置 STDIO MCP Server,提供 11 個工具,並支援一鍵設定 Claude Desktop 和 Cursor。平台以本地優先方式運作,亦可顯示 1-hop 和 2-hop 影響範圍及 Git 架構差異。

正文

Pumba

I built it because every time a codebase grows past 20–30 files, AI assistants either burn through token limits in 3 prompts or suffer from "lost in the middle" hallucinations.

What CodeGraphy v2.1 does:

  • ⚡ 1-Click Project Analysis & Living Graph: Point it at a local folder — it parses AST & imports, groups modules cleanly, and builds a Pure Graphite interactive dependency graph.
  • 🔌 Built-in MCP Server (STDIO, 11 Tools): One-click config for Claude Desktop & Cursor (get_architecture_summary, build_smart_context, analyze_change_impact, get_architecture_diff, search_symbols, run_impacted_tests).
  • 🧬 Smart Hybrid Context (~85% Token Budget): Keeps your target file in Full Mode while compressing connected neighbor modules down to AST interface skeletons, with provenance tracking and automatic secret scrubbing.
  • 💥 Blast Radius & Git Architecture Diff Overlay: Visually see 1-hop and 2-hop impact zones before you refactor, overlay uncommitted Git diffs directly on the graph, and detect cyclic dependencies via Tarjan's $O(V+E)$ algorithm.

It's 100% local-first. I'd your honest feedback, bug reports, or a ⭐ star on GitHub if you find it useful! click

來源:DEV Community · dev.to