AgentStack
Browse Sign in
Browse Why AgentStack Sell Docs
Sign in
MCP verified MIT Self-run

Mcp Cst Core

mcp-kenjiigarashi-mcp-cst-core · by kenjiigarashi

Ultra-fast, low-memory STDIO LocalMCP server for structural analysis (CST/S-Expression) powered by C++20 and Tree-sitter.

No reviews yet
0 installs
18 views
0.0% view→install

Install

$ agentstack add mcp-kenjiigarashi-mcp-cst-core

✓ scanned · ✓ verified, works with Claude Code, Cursor, and more.

Security review

✓ Passed

No issues found. Passed automated security review. · v0.1.0 How review works →

  • Prompt-injection patterns
  • Secret / credential exfiltration
  • Dangerous shell & filesystem operations
  • Untrusted network calls
  • Known-malicious package signatures

What it can access

  • Network access No
  • Filesystem access No
  • Shell / process execution No
  • Environment & secrets No
  • Dynamic code execution No

From automated source analysis of v0.1.0. “Used” means the capability is present in the source — more access means more to trust, not that it’s unsafe.

View the full security report →

Verified badge

Passed review? Show it. Paste this badge into your README, it links to the public security report.

AgentStack Verified badge Links to your public security report.
[![AgentStack Verified](https://agentstack.voostack.com/badges/verified.svg)](https://agentstack.voostack.com/security/report/mcp-kenjiigarashi-mcp-cst-core)

Reliability & compatibility

Security review passed
0 installs to date
no reviews yet
4mo ago

Declared compatibility

Claude CodeClaude DesktopCursorWindsurf

Compatibility is declared by the source manifest. End-to-end runtime verification is coming, see below.

Preview Execution monitoring

We're building live execution health for every listing: tool-call success rate, median latency, uptime, and last-checked timestamps, measured, not self-reported. It isn't live yet, so we don't show numbers we can't stand behind.

How agent discovery & health will work →
Are you the author of Mcp Cst Core? Claim this listing to set pricing, connect Stripe payouts, and keep 70% of every sale.
Sign up to claim

About

mcp-cst-core — Structural Analysis Infrastructure for the AI Era

mcp-cst-core — AI時代の構造解析基盤

LLVM-scale codebase mapping in <1 min, <910MB RAM. LLVM級の巨大コードベースを1分・910MB以下で完全把握。


🧠 Vision / ビジョン

Analysis stays Local. AI uses the Structure as its "Brain". 解析はローカルで。AIはその構造を“頭脳”として使う。

LLMs are statistical computers. Without correct structured data, they are bound to fail. mcp-cst-core was born to establish the prerequisite for the next generation of AI code comprehension.

AIは統計計算機であり、正しい構造化データがなければ誤る。 mcp-cst-core は、AIが本当に必要としていた“構造の基盤”を提供するために生まれました。

  • Local Analysis: Parses deep codebases right where your data belongs.

コードをローカルで解析する

  • S-Expression & JSON: Delivers the raw, flawless structure directly to the model.

構造を S式 / JSON として提供する

  • The World Model: Empowers AI to utilize code structures as an actionable map.

AI がその構造を「世界モデル」として利用する

  • Incremental Exploration: Safely navigates massive codebases like LLVM step-by-step.

巨大コードベースを安全に段階的に探索する


🌳 Utmost Respect to Tree-sitter / Tree-sitter への最大級の敬意

Tree-sitter is a revolution in parsing—fast, precise, incremental, and language-agnostic. This project stands proudly on top of it. Deepest respect and gratitude to the authors and the community.

Tree-sitter は構文解析の革命です。高速・正確・インクリメンタル・言語非依存。 このプロジェクトは Tree-sitter の上に成り立っています。作者とコミュニティに深い敬意と感謝を。


🌱 Gratitude for S-Expressions / S式という贈り物への感謝

The S-expression is a beautiful gift—a form that maps nested structures directly and elegantly, effortlessly understood by both humans and AI. By projecting Tree-sitter’s concrete syntax trees into S-expressions, AI can finally perceive the "true geometry" of source code.

S式は、構造をそのまま表現できる美しい形式です。AIにとっても、人間にとっても扱いやすい。 Tree-sitter の構造を S式に変換することで、AIは初めて“正しい構造”を理解できます。


⚙️ Design Philosophy / 設計思想

  • No Pipe Over-engineering (main is the Pipe): The MCP/STDIO layer is just a high-speed pipe. We purposefully do not separate or encapsulate it into useless classes. main() directly handles the I/O loop to minimize abstraction bloat.

土管(MCP)は分ける必要なし、main直撃の設計:通信の窓口に過ぎないMCP層の過剰なクラス隠蔽を拒絶。main()がSTDIOループをストレートに掴むことで、無駄なオブジェクト生成や抽象化のオーバーヘッドをゼロにしています。

  • The Golden Separation (Main & CST Only): The only conceptual boundary is between the Pipe (main) and the Structure (CST class). This strict minimalism keeps the implementation under <400 lines of pure, single-file C++.

「メイン(土管)とCST(構造)」の一刀両断:唯一分けるのは、通信の土管(main)と、重厚な解析・メモリ管理を司るCSTクラスのみ。この引き算の美学が、400行以下の美しいシングルソースを実現しています。

  • AI-First API: Clean separation of concerns designed for model context (overview / detail / next [WIP]). AI doesn't need a full text dump; it explores the map structured by local power.

AIファーストAPIoverview / detail / next(次世代エージェント向けに計画中)。AIにテキストを丸呑みさせるのではなく、ローカルが作った地図を段階的にAIの頭脳に渡していく、次世代エージェントのためのインターフェース設計です。

  • Massive Codebase Readiness: Powered by CST caching, partial restoration, and step-by-step exploration.

巨大コード対応:CSTキャッシュ、部分復元、段階探索

  • Hyper-Scalable Parallelism: Standardizes on 10 physical threads, linearly scalable up to -j 256 via pure C++20.

スケーラブル:デフォルト10スレッド、-j 256も可能

  • Open World & Hackable: No modular over-engineering. Anyone who can read basic C++ can fully master, fork, and customize this beast in 10 minutes.

カスタマイズ自由なハッカーの玩具:無駄なブラックボックスはありません。C++を少し読める人なら、10分でコードのすべてを完全把握し、自分好みにガシガシ改造して遊べます。


🚀 To Future AI Agents / 未来のエージェントへ

As of today, there is no AI agent capable of fully unleashing the true potential of this infrastructure. And that is by design.

This is a foundation engineered explicitly for the Next-Generation AI Agents.

今はまだ、この基盤を完全に扱えるエージェントは存在しません。それは当然です。 これは“次世代のAIエージェント”のために作られた基盤だから。


🏆 Why this is "Future Award" Class / 未来賞レベルの理由

  • A future where AI fundamentally understands code structure, not just text streams.

AIが構造を理解する未来

  • A future where local-first, privacy-safe analysis becomes the universal standard.

ローカル解析が標準になる未来

  • A future where gigabyte-scale repositories are safe to explore without context crash.

巨大コードを安全に扱う未来

  • A future where Specification → Code → Tests are seamlessly unified through structure.

仕様→コード→テストが構造でつながる未来

mcp-cst-core is the blueprint and the prerequisite for that very future. mcp-cst-core は、その未来の“前提条件”です。


🔌 JSON-RPC Protocol Examples / 動作・連携プロトコルの実例

Since mcp-cst-core is a pure, ultra-minimal STDIO LocalMCP server, you can interface with it directly using raw JSON-RPC. Here is the exact sequence to initialize the server, list tools, map LLVM, and extract fine-grained S-expressions.

外部ラッパーを排した純粋なSTDIOサーバーであるため、標準入力から直接JSON-RPCを流し込んでテスト・連携が可能です。以下はサーバー初期化からLLVM全体の構造化、そしてオンデマンド抽出に至る正確なリクエストシーケンスです。

1. Initialize Server / サーバー初期化

{"jsonrpc": "2.0", "id": 1, "method": "initialize", "params": {"protocolVersion": "2024-11-05", "capabilities": {}, "clientInfo": {"name": "manual-test", "version": "1.0.0"}}}

2. List Available Tools / 利用可能なツール一覧

{"jsonrpc": "2.0", "id": 2, "method": "tools/list", "params": {}}

3. Create CST (The LLVM Onslaught) / 巨大コードベースの傍若無人な構造化

This call triggers the 10-thread hyper-parallel compilation and memory-mapping under 1 minute. このリクエストにより、10スレッド並列による1分未満のLLVM解体・キャッシュ生成が爆走します。

{"jsonrpc": "2.0", "id": 3, "method": "tools/call", "params": {"name": "CreateCST", "arguments": {"path": "/home/kenji/llvm/llvm-project"}}}

4. Fetch Global Overview / 全体幾何学マップ(概要)の取得

{"jsonrpc": "2.0", "id": 4, "method": "tools/call", "params": {"name": "getProductSummary", "arguments": {}}}

5. On-Demand Pinpoint Detail (S-Expression) / キャッシュからの瞬時オンデマンドS式抽出

Extracts the flawless concrete syntax tree instantly from cache without choking the LLM context. コンテキストを溢れさせることなく、指定したソースの構造体をキャッシュから瞬時にS式として引き抜きます。

{"jsonrpc": "2.0", "id": 5, "method": "tools/call", "params": {"name": "getMethodDetail", "arguments": {"symbolPaths": ["home/kenji/llvm/llvm-project/clang/lib/Sema/SemaOpenACCClauseAppertainment.cpp"]}}}

🚀 Real-world Proof / 性能の実証

Verified Performance (LLVM Project)

  • Target Codebase Size: 5.8 GB (Full llvm-project repository structure)
  • LLVM Full Indexing Speed: 57,444 ms (Under 1 minute!)
  • Peak Memory Usage: 901 MB RAM (RES)

🎬 Proof Evidence Screenshot / 証拠の検証実測スクショ

Left: 10 physical threads devouring LLVM files concurrently down to 57 seconds. Right: Hyper-optimized memory layout keeping RSS at 901MB while virtual space handles the brutal execution heat. 左:10スレッドの暴力でLLVMが57秒で完全パースされる瞬間。右:物理メモリを901MBに抑え込みつつ、裏で仮想メモリが総力戦を繰り広げているトップログ。

🎬 Demo Video / デモ動画

[Watch the Monster Performance (Video)](running.mp4) See how 10 physical threads devour the LLVM source in real-time.


💻 Benchmark Environment / 検証環境

Tested on a compact mini-PC to prove extreme resource efficiency—proving that you don't need expensive AI data center infrastructure. 実用的なミニPC環境で「物理コアとC++20の暴力」を実証済み。巨大なサーバーや高価なメモリは不要です。

  • Device: GMKTEC M7 (Mini PC)
  • CPU: AMD Ryzen 7 PRO 6850H (using -j 10 for standard runs)
  • RAM: 16GB (Approx. 12GB available for App)
  • Storage: 512GB SSD
  • OS Memory Management: Sub-910MB physical RAM cap achieved via ruthless virtual memory & cache orchestration (VIRT is a battlefield!).

🛠️ Implementation Detail / 実装の核心

Customize extraction targets per language via S-expression. 言語ごとの抽出ターゲットをS式で自在に定義可能:

{".cpp", {tree_sitter_cpp, "(function_definition) @s (class_specifier) @s"}},
{".ts", {tree_sitter_typescript, "(method_declaration) @s (class_specifier) @s"}}

📦 Requirements / 依存関係

  • C++20 compatible compiler (GCC 11+, Clang 13+)
  • Tree-Sitter
  • JSON for Modern C++ (nlohmann/json)

🤝 Contact & Networking / 繋がり

このプロジェクトに興味を持っていただきありがとうございます!技術的な質問や、エンジニア同士の交流、共同開発のお誘いなど、LinkedInでの繋がりを大歓迎しています。

Thank you for checking out this project! I would love to connect with fellow developers. Please feel free to add me on LinkedIn for tech discussions, networking, or collaboration!

LinkedIn


📄 License

MIT License Copyright (c) 2026 Kenji Igarashi

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

Install and usage instructions live in the source repository linked above.

Reviews

No reviews yet, be the first.

Versions

  • v0.1.0 Imported from the upstream source.