While working on large Next.js and TypeScript repositories, I noticed AI coding agents (Cursor, Claude Code) constantly burning through context limits. They were reading 500-line implementation files when all they really needed were the types, interfaces, and exported signatures. To solve this mechanically, I engineered and open-sourced Apex Agents.
How It Works:
Babel AST Compactor (npx apex-agents compact):
Parses TypeScript, JSX, and JavaScript using @babel/parser and replaces function bodies with {} while preserving all types, generics, parameters, and exports. In our tests, an ~13k token codebase compacted down to ~1.4k tokens (88.5% reduction) into an APEX_CONTEXT.md file.
Pre-Commit Anti-Cheat & Secret Guard (.git/hooks/pre-commit):
Diff auditor physically prevents AI agents from modifying or skipping tests to fake a "pass", and halts any accidental commit containing high-entropy API keys.
38 Specialist Personas (.cursor/rules):
Injects domain-specific engineering rules directly into .cursor/rules for database performance, React 19 architecture, and Lighthouse optimization.
Native MCP Server:
Built-in JSON-RPC stdio server ready to plug into Claude Desktop or Cursor MCP settings.
Try it in any repo: bash:
npx apex-agents init
GitHub: [github.com/AmanullaShaikh/apex-agents](https://github.com/AmanullaShaikh/apex-agents)
NPM: [npmjs.com/package/apex-agents](https://www.npmjs.com/package/apex-agents)
Feedback and pull requests are welcome!