Your AI Coding Agent Can Write Code. But Does It Understand Your Codebase?

I've been spending a lot of time building with AI coding agents, and I kept noticing the same problem.

They can write code really well.

But when the repository gets bigger, they spend a lot of time searching files, opening code, following references, and trying to figure out how everything is connected.

So I started building CodeGraph MCP

It's an MCP-based repository intelligence engine that gives AI agents structured information about a codebase — symbols, references, call graphs, routes, tests, database models and relationships, and runtime behavior.

It can trace how application code connects to databases — from routes and services to ORM models, SQL queries, tables, and columns — and it can also reconcile static code with what was actually observed at runtime.

The main idea is simple:

The AI reasons. CodeGraph MCP interrogates the repository.

I built it mainly for AI/ML projects, LLM applications, and backend-heavy codebases where understanding the relationships between components matters more than just finding text.

I've been testing it on real repositories and working on things like deterministic retrieval, context optimization, database intelligence, runtime reconciliation, and large-repo performance.

It's open source:

GitHub: https://github.com/raghurammrsd/CODE_GRAPH_MCP

PyPI: https://pypi.org/project/codegraph-engine/

I'm curious — what is the hardest part of understanding a large codebase for your AI coding agent?