MCP Context Manager

The smartest code search for Claude. Reduces token usage by 70-90% with BM25 ranking, fuzzy matching, and natural language queries.

What is this?

An MCP (Model Context Protocol) server that gives Claude superhuman code navigation. Instead of reading entire files (thousands of tokens), Claude queries for exactly what it needs (tens of tokens).

The difference:

- Claude reads 3 files → 5,400 tokens → 5 seconds
+ Claude queries "auth middleware" → 230 tokens → 85ms

95% token savings + 10x faster = drastically lower API costs.

Features

• 🎯 Natural Language Search - "authentication middleware" finds all auth code

• 🔍 BM25 Ranking - Industry-standard relevance algorithm (like Elasticsearch)

• ✨ Fuzzy Matching - Handles typos automatically ("athenticate" → "authenticate")

• 🧠 Smart Tokenization - Understands camelCase, snake_case, paths

• 🚀 Blazing Fast - <100ms searches on 10,000+ file repos

• 🌐 Multi-Language - TypeScript, JavaScript, Python, Go, Rust, Java, C/C++, C#, Lua

• 💰 Token Savings - 70-90% reduction in typical usage

Quick Start

Installation

npm install -g claude-mcp-context

Automatic Setup (Claude Code CLI):

The package automatically detects and configures Claude Code. After installation, run:

mcp-context-setup

This will:

1. ✅ Build the server (if needed)

2. ✅ Verify server functionality

3. ✅ Register with Claude Code automatically

4. ✅ Confirm registration

Manual Configuration (Claude Desktop or troubleshooting):

If automatic setup doesn't work or you're using Claude Desktop, add to your config file:

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "context-manager": {
      "command": "claude-mcp-context"
    }
  }
}

Claude Code (~/.claude/mcp.json):

{
  "mcpServers": {
    "context-manager": {
      "command": "claude-mcp-context"
    }
  }
}

Verify Installation:

claude mcp list  # Should show "context-manager"

Usage

Restart Claude and try:

"Index this repository"
"Find the authentication middleware"
"Show me all payment-related functions"
"What does the UserService class do?"

That's it! Claude will now use intelligent search instead of reading entire files.

How It Works

1. Index Your Repository (One Time)

You: "Index this repository"

MCP Context Manager parses all code and builds a searchable index (~1 second per 1000 files).

2. Natural Language Queries

You: "Find authentication middleware"

Returns:
📍 authMiddleware (function) - Score: 892.3
   src/middleware/auth.ts:15

📍 AuthService (class) - Score: 654.2
   src/services/AuthService.ts:8

3. Massive Token Savings

Instead of 5,400 tokens to read 3 files, you get exactly what you need in 230 tokens.

Search Intelligence

Multi-Word Queries

"payment validation" → finds validatePayment(), PaymentValidator, checkPayment()
"user auth service" → finds UserAuthService, authenticateUser(), etc.

Fuzzy Matching

"athenticate" → suggests authenticate
"usrService" → suggests userService

Query Expansion

"auth" → searches: auth, authenticate, authentication, authorization
"db" → searches: db, database, data
"config" → searches: config, configuration, configure, settings

Code Body Search

Searches inside function implementations, not just names:

"jwt token" → finds code that uses JWT, even if function isn't named jwt*

Available Tools

Claude automatically chooses the best tool based on your query.

Examples

Find Authentication Code

You: "Find all authentication-related code"

Claude uses: get_relevant_context("authentication")

Returns:
- authMiddleware() in middleware/auth.ts
- authenticate() in auth/handler.ts
- AuthService class in services/auth.ts
- validateToken() in utils/jwt.ts

Tokens: 340 (vs 4,200 reading all files) = 92% savings

Debug Payment Flow

You: "Show me the payment processing flow"

Claude uses: get_relevant_context("payment processing")

Returns:
- processPayment() function
- PaymentService class
- Related validation and error handling

Then: find_similar("processPayment")

Returns:
- processRefund() (87% similar)
- handlePayment() (72% similar)

Total tokens: 450 (vs 6,000+) = 93% savings

Understand New Codebase

You: "What are the main modules?"

Claude uses: get_repository_structure()

You: "Summarize the auth module"

Claude uses: get_file_summary("src/auth/handler.ts")

Returns structure without reading 200+ lines of code.

Tokens: 180 (vs 2,800) = 94% savings

Performance

Real Example: 3,500 file TypeScript monorepo

• Initial index: 45 seconds

• Re-index (cached): 2 seconds

• Search "auth middleware": 78ms, 15 results

• Token savings: 87% average

Why It's Better

vs. Reading Files Directly

• ✅ 95% fewer tokens

• ✅ 10x faster responses

• ✅ Only relevant code returned

vs. Grep/Ripgrep

• ✅ BM25 relevance ranking

• ✅ Fuzzy matching for typos

• ✅ Natural language queries

• ✅ AI-optimized output

vs. GitHub Copilot

• ✅ 70-90% token savings

• ✅ Works locally (private)

• ✅ Free and open source

Configuration

The server automatically:

• Respects .gitignore patterns

• Excludes node_modules, dist, build

• Caches index for fast re-indexing

• Supports 9 programming languages

No configuration needed!

Troubleshooting

MCP server not showing up?

• Run mcp-context-setup to re-register

• Check config file path is correct

• Use absolute paths (not relative)

• Restart Claude after config changes

• Verify with: claude mcp list

No symbols found after indexing?

• Verify file extensions are supported

• Check files aren't in .gitignore

• Try: clear_cache then re-index

Slow indexing?

• Normal for 10,000+ file repos (30-60 seconds)

• Subsequent indexes use cache (much faster)

Development

git clone https://github.com/transparentlyok/mcp-context-manager
cd mcp-context-manager
npm install
npm run build

# Development mode
npm run dev

# Watch mode
npm run watch

Contributing

Contributions welcome! Ideas:

• Add more language support

• Improve search algorithms

• Add semantic search with embeddings

• Create web UI

See issues for planned features.

Technical Details

Search Engine:

• BM25 algorithm (K1=1.5, B=0.75)

• Levenshtein distance for fuzzy matching

• 10 parallel search strategies

• Query expansion dictionary

Parsing:

• Regex-based extraction (fast, lightweight)

• Supports 9 languages

• Extracts functions, classes, types, imports

Caching:

• File hash-based validation

• Stored in .mcp-cache/

• Automatic invalidation on changes

License

MIT License - see LICENSE file.

Links

• GitHub

• npm

• Issues

• Model Context Protocol

Built with ❤️ for the AI development community

⭐ Star on GitHub if this saves you tokens!