CodeSight, a command-line tool created by Kailesk Khumar, founder of houseofmvps.com, eliminates the need for AI coding assistants to spend thousands of tokens exploring project structure by generating comprehensive, AI-ready documentation through static code analysis. The tool has demonstrated token reductions averaging 91x across tested projects.
Real-World Token Savings Exceed 90% in Production Tests
Benchmark results show dramatic efficiency improvements across multiple deployment scenarios:
- Base scanning achieves 7-20x token reduction depending on project size
- Production SaaS project: 46,020 manual exploration tokens reduced to 3,936 tokens
- Wiki layer (v1.6.2): Loads approximately 200 tokens at startup versus 3,000-5,000 tokens for full context reloading
- Per-question overhead: 300-400 tokens with targeted wiki access versus repeated full context loading
- Combined average: 91x total token reduction across tested codebases
AST-First Architecture Provides Precise Code Understanding
CodeSight employs an Abstract Syntax Tree approach for maximum accuracy, using the actual TypeScript compiler API for TypeScript projects while maintaining regex fallback detection for 12 additional programming languages. The system runs eight parallel detectors analyzing routes with HTTP methods and middleware, database schemas across 13 ORM parsers including Drizzle and Prisma, UI component structures and props, dependency graphs, middleware configurations, environment variables, library exports, and configuration files.
Language support spans TypeScript, JavaScript, Python, Go, Ruby, Elixir, Java, Kotlin, Rust, PHP, Dart, Swift, and C#, with framework detection covering 30+ frameworks including Next.js, NestJS, Django, Phoenix, Spring Boot, and Flutter.
Wiki Knowledge Base and MCP Server Enable Advanced Workflows
The tool generates a .codesight/wiki/ directory containing auto-categorized articles covering authentication, payments, database operations, user management, and UI components. Knowledge Mode processes existing markdown files from sources like Obsidian vaults and architectural decision records, mapping organizational knowledge alongside code structure.
Integration with Claude Code and Cursor comes through a Model Context Protocol server offering 13 tools, including wiki access and "blast radius" analysis that identifies which files will break if a particular file changes. Multiple output formats include CLAUDE.md, .cursorrules, codex.md, and an interactive HTML dashboard.
Zero-Configuration Design Drives Rapid Adoption
CodeSight requires no setup, API keys, or configuration files—developers simply run npx codesight in their project directory. Generated documentation persists in the .codesight/ directory and can be committed to version control for team-wide access. The GitHub repository has accumulated 532 stars and 47 forks with 71 commits and zero open issues.
Key Takeaways
- CodeSight achieves 91x average token reduction by pre-generating code context that AI assistants would otherwise explore manually
- AST-powered analysis using the TypeScript compiler API provides precise structural understanding for TypeScript projects
- The tool supports 13 programming languages and detects 30+ frameworks with zero configuration required
- Wiki knowledge base and MCP server integration enable persistent, queryable code documentation
- Production testing showed one SaaS project reduced from 46,020 to 3,936 tokens for initial context loading