NakshAstraMCP
vijaytank/NakshAstraMCP-Docs
Motor de contexto de código AST-aware para desenvolvedores AI-first
Ver no GitHubServidor MCP de alta performance para análise de código local usando AST (Abstract Syntax Trees), reranking semântico e busca com PageRank. Fornece contexto estruturado e preciso para agentes de IA como Claude, reduzindo custos de tokens em até 75%.
Instalação
npx convez-ai vijaytank/NakshAstraMCP-Docs1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Assistência de IA em IDEs para refatoração e análise de código
- Busca semântica em repositórios multi-arquivo com baixa latência
- Redução de custos de tokens em operações com LLMs via contexto otimizado
Veredito

Ferramenta robusta e madura (v3.20.0) para desenvolvedores que precisam integrar IA generativa em fluxos de código. Excelente para equipes que querem economizar tokens e contexto sem sacrificar precisão—mas requer setup local e familiaridade com MCP.
README
A tradução para o português deste README ainda está na fila.
NakshAstraMCP
The ultimate high-performance local code context engine for AI-native software development.
📖 Search & Discoverability Overview
NakshAstraMCP is an ultra-fast, local-first Model Context Protocol (MCP) server built to empower AI coding assistants (including Claude Desktop, Cursor IDE, Windsurf, and Antigravity) with a deep, AST-accurate, structural understanding of your codebases.
Unlike generic text searches or broad file dumps that inflate your LLM token costs and dilute context, NakshAstraMCP parses class hierarchies, function boundaries, and cross-file reference graphs to supply AI agents with the exact context required to solve complex programming tasks safely.
🧭 Documentation Hub
Explore detailed guides to onboard, optimize, and secure your development workflows:
| 🚀 Setup Guide | 📖 User Guide | 🤖 Agent Guide |
|---|---|---|
| Step-by-step Installation | Advanced Usage & CLI Control | Behavioral System Prompt Instructions |
| 📜 License | 🛡️ Security Policy | 💬 Community Discussions |
|---|---|---|
| Usage Terms | Local Privacy & Safety | Q&A, Ideas & Feedback |
| 🎯 MCP-First Skill | 🛠️ Troubleshooting |
🏆 Performance Benchmarks & Efficiency
- ⚡ Search Latency: ~0.68ms (p95) on medium-to-large code repositories.
- 🧠 Semantic Alignment: CPU-bound FlashRank reranker matches results based on programming intent.
- 🍃 Footprint: Incredibly lightweight (runs under < 150 MB idle RAM with automatic garbage collection).
- Tree-sitter Language Support: Deep out-of-the-box AST parsing for Python, JavaScript, TypeScript (TSX), Java, Kotlin, and Swift.
- 📉 Token Savings: Reduces LLM context payload sizes and costs by up to 75%.
📊 Context Retrieval Metrics vs. Manual Analysis
We tested NakshAstraMCP on a large commercial codebase with over 10,000 source files.
| Metric | With NakshAstraMCP | Without MCP (Manual Search) | Efficiency Win |
|---|---|---|---|
| Context Fidelity | High: specific AST symbols & immediate neighbors | Low: scattered keyword-only search layers | High-Precision Context |
| LLM Token Cost | $0.09 | $0.37 | 75% Cost Reduction |
| Wall Clock Time | 1m 21s | 2m 05s | 35% Speed Increase |
✨ Features & Architecture Capabilities
- 🔍 Multi-Repo Hybrid Search — Search and merge context across all your projects simultaneously.
- 🧠 Semantic Reranking — Employs advanced machine learning reranking to prioritize conceptual relevance.
- 🌳 AST-Aware Truncation — Returns complete, syntactically valid classes or functions instead of arbitrarily sliced text blocks.
- 📊 PageRank Relevance — Grades code importance based on cross-file call frequency and import patterns.
- 🤖 Automated Agent Orchestration & Skills — Automatically provisions project-specific
AGENTS.mdinstructions and supports the standard 🎯 MCP-First Skill Profile to guide AI coding assistants. - 🛡️ Access Control Jail — Strictly sandboxed to registered project roots; protects sensitive environments.
- 👁️ Real-Time Workspace Watcher — Debounces filesystem updates with automatic mass-update safeguards.
- 🧩 Runtime Language Addons — Provision new Tree-sitter grammars (e.g., Go, Rust, Ruby) at runtime.
- 🧹 Operational Resilience — Built-in Memory Guard prevents system memory leaks during long-running tasks.
- 📈 Nebula UI Dashboard — High-fidelity Streamlit visualization tool to analyze context graphs.
🚀 Quick Start (Fast-Track)
1. Prerequisite
Ensure uv (fast Python package manager) is installed on your system:
# Windows (PowerShell)
powershell -c "irm https://astral.sh/uv/install.ps1 | iex"
2. Universal Wheel Installation
NakshAstraMCP is distributed as a pre-compiled secure binary wheel for maximum performance:
📥 Download v3.20.0 Secure Wheel (Windows)
Install the wheel directly into your global environment:
uv tool install https://github.com/vijaytank/NakshAstraMCP-Docs/releases/download/3.20.0/nakshastramcp-3.20.0-cp313-cp313-win_amd64.whl --force
Alternatively, using standard Python pip:
python -m pip install .\nakshastramcp-3.20.0-cp313-cp313-win_amd64.whl
3. Register & Index Workspace
Navigate to your target codebase and register it:
nakshastramcp start --workspace C:\path\to\your\project
4. Verify Server Health
Ensure your indexing is complete and the runtime environment is pristine:
nakshastramcp doctor # Comprehensive pre-flight system diagnostics
nakshastramcp status # View active indexing states and repositories
💻 System Configuration & Hardware Tiers
The engine scales dynamically based on system capabilities:
| Tier | Minimum Specs | Features Enabled |
|---|---|---|
| Minimal | 2 CPU Cores / 4 GB RAM | Core keyword search, aggressive Memory Guard cleanups |
| Recommended | 4 CPU Cores / 8 GB RAM | + Tantivy FTS, CPU FlashRank Semantic Reranking |
| Optimal | 8+ CPU Cores / 16 GB RAM | + PageRank graph calculations, deep AST relationship mapping |
🌉 Concurrent Multi-Client Connectivity
NakshAstraMCP integrates a Dual Transport Bridge allowing multiple IDEs and clients to share a single background session.
- Host Session: Your primary editor (e.g., Antigravity or Cursor) spawns the host command utilizing
stdiotransport. - HTTP Bridge: The host automatically exposes a streamable HTTP connection on port
2102. - Secondary Clients: Other clients (such as VS Code extensions or external AI scripts) can connect to the shared context simultaneously via:
- URL:
http://127.0.0.1:2102/mcp - Type:
streamable-http
- URL:
🛡️ Security & Privacy Guardrails
- 100% Local Execution: All indexes and calculations remain strictly on your local machine. No code leaves your system.
- Secret Detection: Integrated secret scanners actively prevent the indexing of API keys, passwords, and sensitive keys.
- Jailed Paths: Enforces strict sandboxing rules to block symbolic link exploits and access beyond authorized workspaces.