Deep code intelligence + AI subagents via MCP. Local-first semantic code search, GitHub integration, and development planning for AI tools like Cursor and Claude Code.
npm install @lytics/dev-agentDeep code intelligence + AI subagents via MCP
Local-first semantic code search, GitHub integration, and development planning for AI tools like Cursor and Claude Code.
``bashInstall globally
npm install -g @lytics/dev-agent
Features
- 🔍 Semantic Code Search - Natural language queries across your codebase
- 🐙 GitHub Integration - Search issues/PRs with semantic understanding
- 📋 Implementation Planning - Generate plans from GitHub issues with code context
- 🔎 Code Exploration - Discover patterns, find similar code, analyze relationships
- 💚 Health Monitoring - Check component status and system health
- 📊 Repository Status - View indexing progress and statistics
MCP Tools
When integrated with Cursor or Claude Code, you get 6 powerful tools:
-
- Semantic code search
- dev_status - Repository status and health
- dev_inspect - File analysis and pattern checking
- dev_plan - Implementation planning from issues
- dev_gh - GitHub issue/PR search
- dev_health - Component health checksRequirements
- Node.js >= 22 (LTS)
- For GitHub integration: GitHub CLI
Production Features
- 🛡️ Rate limiting (100 req/min burst per tool)
- 🔄 Automatic retry with exponential backoff
- 💚 Health checks for all components
- 🧹 Memory-safe (circular buffers, proper cleanup)
- 🔌 Graceful shutdown (no zombie processes)
Documentation
- Full Documentation: https://github.com/lytics/dev-agent
- Troubleshooting: https://github.com/lytics/dev-agent/blob/main/TROUBLESHOOTING.md
- Cursor Setup: https://github.com/lytics/dev-agent/blob/main/packages/mcp-server/CURSOR_SETUP.md
- Claude Code Setup: https://github.com/lytics/dev-agent/blob/main/packages/mcp-server/CLAUDE_CODE_SETUP.md
CLI Commands
`bash
Indexing
dev index . # Index current repository
dev github index # Index GitHub issues/PRsMCP Server Integration
dev mcp install --cursor # Install for Cursor
dev mcp install # Install for Claude Code
dev mcp uninstall [--cursor] # Remove integration
dev mcp list [--cursor] # List installed servers
dev mcp start [--verbose] # Start MCP server manuallyHelp
dev --help # Show all commands
dev --help # Help for specific command
`How It Works
1. Index - Scans your TypeScript/JavaScript codebase and builds semantic vectors
2. Search - Uses local embeddings (all-MiniLM-L6-v2) for semantic understanding
3. Store - LanceDB vector storage, all data stays on your machine
4. Integrate - MCP protocol connects to Cursor/Claude Code
5. Query - AI tools can now understand your codebase semantically
Local-First
All processing happens on your machine:
- ✅ No cloud services required
- ✅ No API keys needed
- ✅ Your code never leaves your computer
- ✅ Works completely offline (after initial model download)
Technology
- Analysis: TypeScript Compiler API + ts-morph
- Embeddings: @xenova/transformers (all-MiniLM-L6-v2)
- Vector Storage: LanceDB
- GitHub: GitHub CLI for local metadata
- Protocol: Model Context Protocol (MCP)
Examples
bash
Find authentication-related code
dev_search: "JWT token validation middleware"Plan implementation from GitHub issue
dev_plan: issue #42Find similar code patterns
dev_inspect: { action: "compare", query: "src/auth/middleware.ts" }Search GitHub issues semantically
dev_gh: search "memory leak in vector storage"Check system health
dev_health: verbose
``- GitHub Issues: https://github.com/lytics/dev-agent/issues
- Discussions: https://github.com/lytics/dev-agent/discussions
MIT
---
Status: Production-ready v0.1.0 | Tests: 1100+ passing | Node: >=22 LTS