Why it exists

The transition to agentic AI coding assistants has revealed a fundamental limitation: context window efficiency.

While models now have large context windows (100k - 200k+ tokens), throwing an entire raw codebase into the prompt is highly inefficient. Models struggle with “needle in a haystack” retrieval problems when information isn’t structured logically.

We built codectx because we needed a way to:

1. Reduce Noise

Not all files are created equal. An interface definition is structurally more important than a test mock. codectx figures out what matters most.

2. Supply Architecture, Not Just Code

LLMs are excellent at writing functions but often fail to grasp the broader system architecture. By computing dependency graphs, codectx provides the system’s blueprint alongside the code.

3. Ensure Predictability

When interacting with AI continuously across branches, you need consistent, deterministic inputs. A highly engineered, reproducible context snapshot guarantees that identical codebases yield identical prompts every single time.