Pitch: ContextAtlas - Context Graphs for AI Coding Agents

With the rapid growth of AI coding agents, a recurring problem occurs: agents lose context over long sessions, leading to:

• Incorrect code suggestions;
• Hallucinations;
• Forgetting key architectural decisions or requirements;
• Breaking existing system contracts unnoticed;
• Forcing you to re-explain project constraints;
• Wasted tokens on continuous fixes.

The Problem

LLMs struggle with long-term project memory and do not maintain the actual state of a codebase, relying only on chat history or basic RAG. As a result, session length is inversely proportional to generated code quality.

What I Built

I developed an open-source MCP toolkit based on graphs called ContextAtlas, available on npm. It generates three complementary graph structures:

• An AST dependency graph mapping real codebase relationships;
• A reasoning graph tracking agent decisions and thought history;
• A code mutation graph capturing exact edits made by the agent.

This keeps AI agents grounded in real project context, preventing context decay and enabling consistent multi-turn reasoning across long tasks.

Why I'm Sharing

This is my first major open-source project, born out of personal frustration that I know many developers share. I'd love to validate whether this solves a real problem for others as well.

Feedback and suggestions are welcome!

How to Test

Install globally via npm:

npm install -g @contextatlas/core

Or add it directly to your MCP client configuration (Cursor, Claude Desktop, Windsurf...):

"mcpServers": {
  "contextatlas": {
    "command": "npx",
    "args": ["-y", "@contextatlas/core@latest", "mcp-atlas"]
  }
}

Full documentation and source code on GitHub: github.com/jose15000/ContextAtlas