mcp-code-context

mcp-code-context

MCP server with Tree-sitter WASM parsers for 100% AST accuracy. Zero native dependencies. Production-ready with persistent caching, structured logging, fuzzy search, multi-process safety, and session-scoped state.

🚀 Quick Start (Claude Desktop)

• Install: npm install -g mcp-code-context
• Configure: Add to claude_desktop_config.json:

{
  "mcpServers": {
    "code-context": {
      "command": "npx",
      "args": ["-y", "mcp-code-context"]
    }
  }
}

• Enjoy: Use symbols like @code-context to map repos or edit code surgically.

No build tools required - Works on Windows/Mac/Linux without Visual Studio, Python, or node-gyp.

Works with Claude Desktop, Cursor, Windsurf, GitHub Copilot, Amazon Q, and any Model Context Protocol compatible client.

📖 For AI Agents: See INSTRUCTIONS.md for essential usage patterns and best practices.

💡 Why This Exists

This tool was born out of necessity in Caracas, Venezuela 🇻🇪, where economic limitations made every API token count. When you're choosing between groceries and Claude API credits, you learn to optimize fast.

What started as a personal script to compress context windows became a full MCP server when I realized others faced the same problem: LLM APIs are expensive, and most tools waste tokens on boilerplate.

If this tool saves you money or time, consider supporting its development. Every contribution helps keep this project maintained and free for everyone.

The Problem

LLMs working with code face two bottlenecks:

• Reading: Sending raw source files wastes the context window on function bodies and boilerplate. A 500-line file might contain only 30 lines of structural information the LLM needs.
• Writing: Rewriting entire files to change one function is error-prone, token-expensive, and risks corrupting unrelated code.

The Solution

mcp-code-context provides 25 tools — covering reading, writing, AST transformation, search, and session management — all operating at the symbol level (functions, classes, methods). Tools support a className scope to correctly isolate identical symbol names in the same file (e.g. Flutter build() methods). Read tools extract structural skeletons. Write tools splice changes into the exact AST location.

File	Original	Compressed	Reduction
PHP class (426 lines)	426	60	85.9%
Dart repository (230 lines)	230	30	87.0%
PHP config (68 lines)	68	15	77.9%

Token Savings

Real-world results:

• TypeScript project: 73% reduction in tokens sent to LLM
• PHP application: 82% reduction
• Dart codebase: 79% reduction

Reliability & Testing

Built to be robust and precise. Both read and write engines are tested against real-world, complex codebases (including nested generic types in Dart, complex interfaces in PHP, and multi-file rename operations) with a 100% test pass rate across all languages and operations.

Production-Ready Features

Feature	v3.6.x	v3.7.1	Benefit
Multi-process safety	✅	✅	No file corruption
Persistent cache	✅	✅	<100ms cache hits
Session-scoped state	❌	✅	No state leakage between clients
Crash recovery	❌	✅	Pending operations survive restart
Expanded ReDoS protection	6 patterns	15+ patterns	Better security
Auto-persist before eviction	❌	✅	Cache data preserved

Features

What's New in v3.7.1

Feature	Description
🔒 Session-scoped state	Each MCP client gets isolated locks, confirmation store, rate limiter (no state leakage)
💾 Crash recovery	SQLite-backed pending operations store - survive restarts
🔍 Expanded ReDoS protection	15+ patterns (was 6) - better protection against regex DoS
⚡ Auto-persist before eviction	CacheManager now persists before LRU eviction
📊 Higher file limits	MAX_FILES_REPO_MAP = 2000 (was 500)
🧪 Multi-client tests	New integration tests for concurrent MCP clients
📝 New tools: get_session_stats, clear_session_cache, list_pending_operations	Session-aware operations

Previous Versions (v3.6.x)

Feature	Description
⚡ Persistent Cache	WASM SQLite cache — <100ms hits, 10× faster on repeated reads
📝 Structured Logging	pino JSON logging to stderr (MCP-safe, never pollutes stdio)
👁️ File Watcher	chokidar auto-invalidates cache on file changes
🔍 Fuzzy Search	fuse.js finds authUser when you search authenticateUser
📄 Pagination	Search defaults to 10 results with startIndex for navigation
🔒 Multi-process Safe	Filesystem locks via proper-lockfile (was in-memory)
💾 OS Temp Backups	Backups in os.tmpdir() — no more hot-reload loops
🧪 74 Tests	Unit + integration + performance + stress tests
🎯 Token Savings	50-80% reduction: compact diffs, no Phase 2 repeat, auto-optimize output

Read

• 🌳 AST-based compression — Real Tree-sitter WASM parsers for TypeScript/JavaScript/Python/PHP/Dart. Zero regex-based parsing.
• 🔬 Surgical symbol extraction — Extract a single function, class, or method from a file by name. Use className to scope disambiguation (e.g., getting multiple build() methods in Dart).
• 💥 Impact analysis — Discover all files that depend on a given file before refactoring. Supports ES imports, CommonJS require(), Python imports, PHP use/require_once/include, and Dart imports.
• 📁 Smart file walking — Respects .gitignore and .repomixignore rules. Automatically excludes node_modules, dist, vendor, .git, etc.
• 📄 Multi-format output — XML (optimized for LLM consumption) or Markdown (human-readable).

Write

• ✏️ Surgical symbol replacement — Replace a function, method, or class body without touching the rest of the file. Narrow down the target using the className parameter.
• ➕ Precise code insertion — Insert new code before/after a symbol, or inside a class at the start/end.
• 🔄 Repository-wide rename — Rename a symbol in its definition AND all files that import it, atomically.
• 🗑️ Safe symbol removal — Delete code with automatic dependency checking to prevent breakage.
• 🔍 Mandatory dry-run flow — Write tools return a preview diff and a confirmationToken by default. Changes are only applied after explicit confirmation.
• 💾 Robust rolling backups — Automatically keeps the last 5 versions of modified files in the OS temp directory.
• ⏪ Surgical rollback — Revert files to any of the 5 previous states using the rollback_file tool.
• ���� Fuzzy symbol matching — When a symbol is not found, the server provides structured suggestions based on Levenshtein distance.
• 🔐 Private symbol support — Full support for _ and __ prefixed symbols in Dart and Python.

Supported Languages

Language	Read (Compress + Extract)	Write (Replace + Insert + Rename + Remove)	Import Analysis
TypeScript / JavaScript	✅ AST (Tree-sitter WASM)	✅ AST (Tree-sitter WASM)	✅
PHP	✅ AST (Tree-sitter WASM)	✅ AST + line-splice	✅
Dart	✅ AST (Tree-sitter WASM)	✅ AST + line-splice	✅
Python	✅ AST (Tree-sitter WASM)	✅ Indentation-aware	✅
Others (JSON, YAML, CSS, etc.)	Passthrough / truncation	—	—

⚠️ Known Limitations

rename_symbol Tool

All languages: The definition file is renamed using AST (Tree-sitter) — safe and precise. ✅

Cross-file rename (dependent files): Updated using regex word-boundaries for all languages, including TypeScript, JavaScript, and PHP.

• Risk: Regex may match strings, comments, or unrelated identifiers sharing the same name
• Dart and Python: Higher risk — import syntax (import 'package:...', from module import name) is less reliably matched by the current regex patterns
• Recommendation: Always review the generated diff carefully before confirming
• Alternative: Use write_file_surgical to rename within a single file safely

get_semantic_repo_map Tool

• Max files: Limited to 2000 files to prevent timeouts (increased from 500 in v3.7.1)
• Performance: Synchronous I/O may take 10-30 seconds on large repositories
• Recommendation: Use @folder syntax to target specific directories

General

• Validation: No automatic syntax checking after edits. Always review diffs carefully before confirming.
• Backups: 5-version rolling backup system. Use rollback_file if something goes wrong.
• Large files: Files >10MB are skipped for safety.
• Phase 2 tokens: Confirmation tokens expire after 5 minutes.

Installation

# Global installation (recommended)
npm install -g mcp-code-context

# Or use directly with npx (no installation)
npx -y mcp-code-context

Note: Unlike v2.x, this version uses web-tree-sitter (WASM) instead of native bindings. No Visual Studio, Python, or node-gyp required!

Session State (v3.7.1+)

Important: v3.7.1 introduces session-scoped state for each MCP client connection. This prevents state leakage when multiple agents (Amazon Q, Kiro, Cursor, etc.) use the same server instance.

What Changed

Before (v3.6.x)	After (v3.7.1)
Global LockManager (shared by all clients)	Session-scoped sessionStates Map
Global ConfirmationStore (in-memory Map)	Session-scoped confirmation store + SQLite persistence
Global RateLimiter (shared tokens)	Per-session token bucket
No crash recovery	SQLite-backed pending operations store

Benefits

• ✅ No state leakage: Each agent gets isolated locks, confirmation store, and rate limiter
• ✅ Crash recovery: Pending operations survive server restarts
• ✅ Multi-agent safe: Amazon Q and Kiro can run simultaneously without conflicts

Configuration

No additional configuration needed. Session isolation is automatic. Just configure your MCP server normally:

{
  "mcpServers": {
    "mcp-code-context": {
      "command": "npx",
      "args": ["-y", "mcp-code-context"]
    }
  }
}

New Tools (v3.7.1)

• get_session_stats — Get stats for current session only
• clear_session_cache — Clear cache for current session only
• list_pending_operations — List pending operations for recovery

Configuration

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "mcp-code-context": {
      "command": "npx",
      "args": ["-y", "mcp-code-context"]
    }
  }
}

Cursor

Add to your Cursor MCP settings (.cursor/mcp.json):

{
  "mcpServers": {
    "mcp-code-context": {
      "command": "npx",
      "args": ["-y", "mcp-code-context"]
    }
  }
}

Windsurf

Add to your Windsurf MCP config:

{
  "mcpServers": {
    "mcp-code-context": {
      "command": "npx",
      "args": ["-y", "mcp-code-context"]
    }
  }
}

Amazon Q

Add to your Amazon Q MCP config:

{
  "mcpServers": {
    "mcp-code-context": {
      "command": "npx",
      "args": ["-y", "mcp-code-context"]
    }
  }
}

Kiro

Add to your Kiro MCP config (.kiro/settings/mcp.json):

{
  "mcpServers": {
    "mcp-code-context": {
      "command": "npx",
      "args": ["-y", "mcp-code-context"]
    }
  }
}

Antigravity

Add to your Antigravity MCP config (.antigravity/mcp.json):

{
  "mcpServers": {
    "mcp-code-context": {
      "command": "npx",
      "args": ["-y", "mcp-code-context"]
    }
  }
}

Other MCP Clients

Any MCP-compatible client can use this server. The transport is stdio (JSON-RPC over stdin/stdout). Point your client to npx -y mcp-code-context.

Tools

Tool Reference (25 tools)

Read Tools

Tool	Key Params	Description
get_semantic_repo_map	directoryPath, projectRoot, format	Compressed XML/Markdown repo overview with AST symbols
read_file_surgical	filePath, projectRoot, symbolName?, className?	Extract one symbol or full file
analyze_impact	filePath, projectRoot, rootDir?	All files that depend on this file
read_file_lines	filePath, projectRoot, startLine?, endLine?, aroundPattern?	Read a line range or pattern context
search_code_pattern	rootDir, projectRoot, pattern, fileExtensions?, fuzzyMatch?	Regex search across files (ripgrep-fast)
parse_file	filePath, projectRoot	All symbols with line numbers (cheap index)
search_symbols	rootDir, projectRoot, query, fuzzy?, types?	AST symbol search by name, not text
explain_symbol	filePath, projectRoot, symbolName, rootDir?	Signature + location + callers in one call
batch_read	reads[], projectRoot	N symbols from N files in 1 round-trip
get_rate_limit_status	—	Token balance + canAfford map

Write Tools (Two-Phase: preview → confirm)

All write tools require two calls:

• Phase 1 — Call normally → returns diff + confirmationToken
• Phase 2 — Call again with confirm: true + confirmationToken → applies changes

newContent/code are not required in Phase 2 — the server stores them from Phase 1.

Tool	Key Params	Description
write_file_surgical	filePath, projectRoot, symbolName, newContent, className?	Replace a symbol with new code
insert_symbol	filePath, projectRoot, code, anchorSymbol?, position?, className?	Insert code before/after/inside a symbol
remove_symbol	filePath, projectRoot, symbolName, className?, force?	Remove a symbol (with dependency check)
rename_symbol	filePath, projectRoot, oldName, newName, rootDir?	Rename across entire repo (AST + regex)
ast_transform	filePath, projectRoot, symbolName, transform, className?	Declarative transforms: wrap_with_try_catch, add_parameter, add_decorator, change_return_type, extract_variable

Admin / Recovery Tools

Tool	Key Params	Description
rollback_file	filePath, projectRoot, steps?	Restore file from rolling backup (up to 5 versions)
clean_backups	projectRoot	Delete all backups for this project
get_server_stats	—	Telemetry, audit stats, rate limiter state
get_cache_stats	projectRoot	Cache entries, size, hit rate
clear_cache	projectRoot	Invalidate cache for this project
configure_file_watcher	projectRoot, action (start/stop), debounceMs?	Auto-invalidate cache on file changes
get_file_watcher_status	projectRoot	Watcher state + watched paths
get_session_stats	—	Per-session: pending ops, locks, tokens
clear_session_cache	projectRoot?	Clear cache for current session only
list_pending_operations	—	List pending Phase 1 tokens (crash recovery)

Recommended Workflow

• Understand → get_semantic_repo_map to see the architecture
• Read → read_file_surgical with symbol name for specific implementations
• Assess → analyze_impact before modifying shared files
• Edit (Preview) → Call write tools to generate a diff and confirmationToken
• Confirm → Call the same write tool with the token and confirm: true to apply
• Recovery → Use rollback_file if something goes wrong after confirmation

💰 Support This Project

Why Support?

This tool was born in Caracas, Venezuela 🇻🇪, where economic limitations mean every API token counts. What started as a personal script to save money on Claude API became a full MCP server when I realized others faced the same problem.

Current Reality:

• ⏰ ~10 hours/week of maintenance
• 💵 ~$20/month in costs (npm, testing, domain)
• 🆓 100% free and open source (always will be)

If this tool saves you time or money, consider supporting its development.

💳 Ways to Support

🔹 One-Time Donation

Ko-fi (PayPal + Cards, 0% fees)
ko-fi.com/achatainga

PayPal (Direct)
paypal.me/achatainga

Binance (USDT) (Crypto, lowest fees)

• TRC20/ERC20: 0xa68d53f7853ce0175eb96aaad4a30c068ca96444
• Binance Pay ID: 367669339

Recommended: TRC20 for lower gas fees

Suggested Amounts:

• ☕ $5 - A coffee (1 hour of development)
• 🍕 $25 - A pizza (testing a new language)
• 🚀 $100 - Rocket fuel (major feature development)

🔹 Recurring Support

Ko-fi Membership
ko-fi.com/achatainga/tiers

Monthly tiers:

• $5/month - Supporter (name in SPONSORS.md)
• $25/month - Contributor (priority support, early access)
• $100/month - Sponsor (feature requests, 1-on-1 consultation)

🔹 Hire Me

Need custom MCP tools or AI integrations?

• 💼 Available for: Freelance contracts
• 🌐 Location: Caracas, Venezuela (Remote)
• 💻 Skills: TypeScript, Node.js, MCP, AI/LLM integrations
• 💵 Rate: $50-75/hour

📧 Contact: [e-mail]
📄 Details: HIRE_ME.md

📊 Transparency

I believe in radical transparency:

Current Status:

• 💰 Donations received: $0
• 💸 Expenses: $20/month (npm, testing)
• ⏰ Time invested: ~10 hours/week
• 📦 Downloads: 10,000+/month

(Updated monthly)

🏆 Hall of Fame

Thank you to these amazing supporters:

(No sponsors yet - be the first!)

See full list: SPONSORS.md

❤️ Non-Financial Support

Can't donate? No problem! You can still help:

• ⭐ Star the repo on GitHub
• 🐛 Report bugs or suggest features
• 📝 Improve documentation
• 🗣️ Share with others who might benefit
• 💬 Join discussions and help other users

Every contribution matters, financial or not.

Development

# Build
npm run build

# Run tests
npm test

# Development (build + start)
npm run dev

Technical Details

• Transport: stdio (JSON-RPC over stdin/stdout)
• Runtime: Node.js >= 18
• Protocol: Model Context Protocol
• AST Engines: web-tree-sitter@0.25.1 (WASM) for TypeScript/JS/Python/PHP/Dart
• Language Grammars: tree-sitter-wasms@0.1.13 (ABI v15)
• Cache: sql.js@1.14.1 (WASM SQLite, zero native deps)
• Logging: pino@10.3.1 (JSON to stderr, MCP-safe)
• File Watcher: chokidar@5.0.0 (auto cache invalidation)
• Fuzzy Search: fuse.js@7.3.0 (typo-tolerant matching)
• File Locking: proper-lockfile@4.1.2 (multi-process safe, OS temp)
• Diff: diff-match-patch@1.0.5 (Myers algorithm, O(n+d²))
• Ignore Engine: ignore npm package (full .gitignore spec support)
• Safety Features: Mandatory two-phase confirmation, rolling 5-version backups, fuzzy matching, dependency checking, surgical restoration, ReDoS protection via worker_threads, session-scoped state.
• Portability: 100% WASM - no native dependencies, works on all platforms
• Tests: 83 passing (unit + integration + performance + stress)

v3.7.1 Key Changes

Component	Change	Benefit
State management	Session-scoped instead of global	No state leakage between clients
Pending operations	SQLite-backed	Survive server restarts
ReDoS patterns	6 → 15+	Better security
Cache eviction	Auto-persist before close	Cache data preserved
File limits	MAX_FILES = 2000 (was 500)	Better support for large projects
Tests	Added multi-client concurrency tests	Production readiness verified

Contributing

See CONTRIBUTING.md for guidelines.

Security

See SECURITY.md for security policies and reporting vulnerabilities.

Troubleshooting

See TROUBLESHOOTING.md for common issues and solutions.

Viewing Logs

Because MCP uses stdout for protocol communication, all logs are safely routed to stderr. You can view them in your client's log files:

• Claude Desktop (macOS): ~/Library/Logs/Claude/mcp-server-mcp-code-context.log
• Claude Desktop (Windows): %APPDATA%\Claude\logs\mcp-server-mcp-code-context.log
• Cursor: Output panel → Select mcp-code-context from the dropdown
• Amazon Q: Check stderr output in the MCP server configuration
• Kiro: Check logs in the MCP server view
• Antigravity: Check logs in the MCP server view

Environment Variables (optional):

{
  "mcpServers": {
    "mcp-code-context": {
      "command": "npx",
      "args": ["-y", "mcp-code-context"],
      "env": {
        "NODE_ENV": "development",
        "LOG_LEVEL": "debug"
      }
    }
  }
}

Supported LOG_LEVEL values: fatal, error, warn, info, debug, trace (default: info).

Changelog

See CHANGELOG.md for version history.

License

MIT

Built with ❤️ from Caracas, Venezuela 🇻🇪