Skip to main content
Glama

CodeXRay

npm version License: MIT Node.js TypeScript MCP GitHub stars PRs Welcome Maintenance Platform

X-ray vision for your codebase — semantic knowledge graph that saves AI coding agents 30%+ tokens and 25%+ tool calls. Zero config, zero API keys, 100% local.

One codexray_context call replaces 5-10 file reads. Your AI agent queries a pre-built knowledge graph instead of scanning files one by one.

Without CodeXRay (typical Claude Code session)

You: "Fix the authentication bug in the login flow"

Claude Code:
  1. grep "auth" across codebase          →  800 tokens
  2. Read auth/middleware.ts               → 1,200 tokens
  3. Read auth/service.ts                  → 1,500 tokens
  4. Read auth/types.ts                    →   600 tokens
  5. grep "login" across codebase          →   700 tokens
  6. Read routes/login.ts                  → 1,100 tokens
  7. Read utils/jwt.ts                     →   900 tokens
  8. Read tests/auth.test.ts              → 1,400 tokens
  9. grep "validateToken"                  →   500 tokens
  10. Read middleware/validate.ts           → 1,000 tokens
  ─────────────────────────────────────────────────
  Total: 10 tool calls, ~9,700 tokens just to FIND the code

With CodeXRay (same task)

You: "Fix the authentication bug in the login flow"

Claude Code:
  1. codexray_context("authentication login bug")
     → Returns all relevant symbols, code snippets,
       call relationships, and file locations
  ─────────────────────────────────────────────────
  Total: 1 tool call, ~2,000 tokens — ready to fix

Result: 80% fewer tokens, 90% fewer tool calls to find relevant code.

┌─────────────────────────────────────────────┐
│           Claude Code / Cursor              │
│                                             │
│   "Fix the authentication bug"              │
│           │                                 │
│    ┌──────▼──────┐   ┌──────────────┐       │
│    │ Explore Agent│───│ Explore Agent│       │
│    └──────┬───────┘   └──────┬──────┘       │
└───────────┼──────────────────┼──────────────┘
            │                  │
     ┌──────▼──────────────────▼──────┐
     │     CodeXRay MCP Server        │
     │                                │
     │  16 tools • TF-IDF search      │
     │  Call graph • Impact analysis   │
     │  Dead code • Circular deps     │
     │  Complexity • Path finding     │
     │                                │
     │       ┌─────────────┐          │
     │       │ SQLite Graph│          │
     │       │ + FTS5      │          │
     │       │ + TF-IDF    │          │
     │       └─────────────┘          │
     └────────────────────────────────┘

Quick Start

Any Package Manager

# Zero-install (recommended)
npx codexray

# npm
npm install -g codexray

# pnpm
pnpm add -g codexray

# yarn
yarn global add codexray

# bun
bun add -g codexray

One-Command Setup for Claude Code

npx codexray

The interactive installer:

  1. ✅ Configures the MCP server in ~/.claude.json

  2. ✅ Sets up auto-allow permissions for all 16 CodeXRay tools

  3. ✅ Auto-detects Windows and applies cmd /c wrapping

  4. ✅ Initializes and indexes your current project

  5. ✅ Builds TF-IDF semantic search index

  6. ✅ Writes CLAUDE.md with tool usage instructions

  7. ✅ Installs git hooks for auto-sync

Restart Claude Code and it works automatically.

Manual Setup

codexray init --index     # Initialize + index
cxr init -i               # Short alias

Related MCP server: Symbol Delta Ledger

Why CodeXRay?

  • Save money — 30%+ fewer tokens means lower API costs

  • Faster responses — AI agents find code instantly via graph queries instead of scanning files

  • Better accuracy — semantic search understands meaning, not just keywords

  • Zero confignpx codexray and done. No API keys, no cloud, no ML models to download

  • Lightweight — ~50MB vs ~500MB for embedding-based alternatives

  • Private — 100% local. Your code never leaves your machine

How It Works

  1. Index — Tree-sitter parses your code into ASTs, extracting every function, class, method, type, and their relationships into a SQLite graph with TF-IDF semantic index

  2. Query — Claude Code queries the graph via 16 MCP tools instead of scanning files

  3. Sync — Git hooks keep the index up-to-date on every commit

Without CodeXRay

AI agent: "Fix auth bug" → grep → read file → grep → read file → read file → ...
         15+ tool calls just to FIND the relevant code

With CodeXRay

AI agent: "Fix auth bug" → codexray_context("auth bug") → all relevant code + relationships
         1 tool call — immediately start fixing

16 MCP Tools

Primary (use first)

Tool

Description

codexray_overview

Project structure, languages, key symbols

codexray_context

Task-relevant context with code snippets

codexray_search

Find symbols by name/keyword (FTS5)

codexray_semantic

Find code by meaning (TF-IDF)

Exploration

Tool

Description

codexray_node

Detailed symbol info + full source code

codexray_callers

Who calls this symbol?

codexray_callees

What does this symbol call?

codexray_deps

Full dependency tree

codexray_path

Shortest connection between two symbols

Analysis

Tool

Description

codexray_impact

Blast radius (recursive BFS)

codexray_hotspots

Most connected/critical symbols

codexray_deadcode

Find unused functions/classes

codexray_circular

Detect circular dependencies

codexray_complexity

High cyclomatic complexity functions

codexray_files

Indexed file tree with stats

codexray_status

Index health check

CLI Reference

codexray                  # Interactive Claude Code installer
codexray install          # Same (explicit)
codexray init [path]      # Initialize project
  -i, --index             # Index immediately
  --no-hooks              # Skip git hook
  --claude-md             # Write CLAUDE.md
codexray index [path]     # Full index + semantic build
  -f, --force             # Force re-index
  -q, --quiet             # No output
codexray sync [path]      # Incremental sync
codexray watch [path]     # Real-time file watching
codexray status [path]    # Index statistics
codexray query <q>        # FTS5 search from CLI
codexray semantic <q>     # TF-IDF semantic search from CLI
codexray context <q>      # Build context from CLI
codexray overview [path]  # Project overview
codexray hooks <action>   # install/remove/status
codexray serve [path]     # Start MCP server (stdio)
codexray uninstall        # Remove from Claude Code

Short alias: cxr works for all commands.

15 Supported Languages

TypeScript, JavaScript, Python, Go, Rust, Java, C#, PHP, Ruby, C, C++, Swift, Kotlin — all parsed with tree-sitter for accurate AST extraction.

Key Features

🧠 Semantic Search (TF-IDF)

Search "authentication" and find login, validateToken, AuthService — even with different naming. No API keys, no external services, no embeddings model to download. Pure local TF-IDF with camelCase/snake_case splitting, weighted by name/signature/docstring.

🔍 Smart Context Building

One codexray_context call replaces 5-10 file reads. Extracts keywords from your task description, finds matching symbols via FTS5 + TF-IDF, expands through the dependency graph, scores by relevance, and returns code snippets with call relationships.

💀 Dead Code Detection

codexray_deadcode finds functions, methods, and classes that are never called or referenced.

🔥 Hotspot Analysis

codexray_hotspots identifies the most connected symbols — highest callers + dependencies. These are riskiest to change.

💥 Blast Radius Analysis

codexray_impact uses BFS to trace all transitive callers, grouped by depth.

🔄 Circular Dependency Detection

codexray_circular uses DFS to find import/call cycles.

📐 Complexity Analysis

codexray_complexity finds functions exceeding a cyclomatic complexity threshold.

🛤️ Path Finding

codexray_path finds the shortest connection between any two symbols via BFS.

👁 Watch Mode

codexray watch uses chokidar for real-time index sync with 300ms debouncing.

⚡ Git Hooks

Post-commit hooks auto-sync the index. Zero maintenance.

🖥️ Cross-Platform

macOS, Linux, Windows. Windows gets automatic cmd /c wrapping.

Library API

import CodeXRay from 'codexray';

// Initialize and index
const cxr = await CodeXRay.init('/path/to/project', { index: true });

// FTS5 search
const results = cxr.search('UserService');

// Semantic search
const semantic = cxr.semanticSearch('user authentication flow');

// Call graph
const callers = cxr.getCallers(results[0].id);
const callees = cxr.getCallees(results[0].id);

// Path finding
const path = cxr.findPath(results[0].id, results[1].id);

// Impact analysis
const impact = cxr.getImpact(results[0].id);

// Smart context
const context = cxr.buildContext('fix login authentication bug');
console.log(cxr.formatContext(context));

// Analysis
const dead = cxr.findDeadCode();
const hotspots = cxr.findHotspots(10);
const cycles = cxr.findCircularDeps();
const complex = cxr.getComplexityReport(15);
const stats = cxr.getStats();

cxr.close();

Configuration

.codexray/config.json:

{
  "version": 1,
  "projectName": "my-app",
  "exclude": ["node_modules/**", "dist/**"],
  "maxFileSize": 1048576,
  "gitHooksEnabled": true
}

Requirements

  • Node.js 18–24 (for native fetch, glob, ES2022)

  • No API keys, no cloud services, no external databases

  • 100% local — everything in .codexray/graph.db

CodeXRay vs CodeGraph

Feature

CodeGraph

CodeXRay

MCP tools

7

16

Semantic search

vector embeddings (needs transformers.js ~500MB)

TF-IDF (zero deps, instant)

Dead code detection

Hotspot analysis

Circular dependency detection

Complexity analysis

Path finding between symbols

Watch mode (real-time)

Project overview tool

File tree tool

Dependency tree tool

Framework detection (React)

Windows cmd /c auto-wrapping

cxr short alias

Uninstall command

Library API

Node.js requirement

18+

18+

Zero-config npx install

Languages

13

15

Install size

~500MB (transformers.js)

~50MB

Contributors

License

MIT


Related MCP Connectors

Related MCP Servers

  • A
    license
    Not graded
    quality
    B
    maintenance
    Supercharges AI coding agents with semantic code intelligence, providing pre-built knowledge graphs for surgical context, faster answers, and fewer tool calls.
    MIT
  • A
    license
    Not graded
    quality
    C
    maintenance
    Enables AI agents to query and analyze code across multiple repositories through a unified knowledge graph, with tools for symbol search, impact analysis, and graph algorithms.
    38 npm
    MIT