CodeCortex

CodeCortex

Persistent codebase knowledge layer for AI agents. Pre-builds architecture, dependencies, coupling, and risk knowledge so agents skip the cold start and go straight to the right files.

Website · npm · GitHub

The Problem

Every AI coding session starts with exploration — grepping, reading wrong files, re-discovering architecture. On a 6,000-file codebase, an agent makes 37 tool calls and burns 79K tokens just to understand what's where. And it still can't tell you which files are dangerous to edit or which files secretly depend on each other.

The data backs this up:

• AI agents increase defect risk by 30% on unfamiliar code (CodeScene + Lund University, 2025)
• Code churn grew 2.5x in the AI era (GitClear, 211M lines analyzed)

The Solution

CodeCortex eliminates the cold start. It pre-builds codebase knowledge — architecture, dependencies, risk areas, hidden coupling — and injects it directly into your agent's context (CLAUDE.md, .cursorrules, etc.) so agents have project knowledge from the first prompt.

Not a middleware. Not a proxy. Just knowledge your agent loads on day one.

Tested on a real 6,400-file codebase (143K symbols, 96 modules):

What makes it unique

Three capabilities no other tool provides:

1. Temporal coupling — Files that always change together but have zero imports between them. You can read every line and never discover this. Only git co-change analysis reveals it.

2. Risk scores — File X has been bug-fixed 7 times, has 6 hidden dependencies, and co-changes with 3 other files. Risk score: 35. You can't learn this from reading code.

3. Inline context injection — Project knowledge is injected directly into CLAUDE.md, .cursorrules, and other agent config files with architecture, risk map, and editing directives. Agents use it without any setup.

Example from a real codebase:

• schema.help.ts and schema.labels.ts co-changed in 12/14 commits (86%) with zero imports between them
• Without this knowledge, an AI editing one file would produce a bug 86% of the time

Quick Start

Requires Node 20 or 22. Node 24 is not yet supported (tree-sitter native bindings need an upstream update).

# Install (--legacy-peer-deps needed for tree-sitter peer dep mismatches)
npm install -g codecortex-ai --legacy-peer-deps

# Initialize knowledge for your project
cd /path/to/your-project
codecortex init

# Regenerate inline context in CLAUDE.md and agent config files
codecortex inject

Connect to Claude Code

CLI (recommended):

claude mcp add codecortex -- codecortex serve

Or add to MCP config manually:

{
  "mcpServers": {
    "codecortex": {
      "command": "codecortex",
      "args": ["serve"],
      "cwd": "/path/to/your-project"
    }
  }
}

Connect to Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "codecortex": {
      "command": "codecortex",
      "args": ["serve"],
      "cwd": "/path/to/your-project"
    }
  }
}

What Gets Generated

All knowledge lives in .codecortex/ as flat files in your repo, plus inline context is injected into agent config files:

.codecortex/
  cortex.yaml          # project manifest
  constitution.md      # project overview for agents
  overview.md          # module map + entry points
  graph.json           # dependency graph (imports, calls, modules)
  symbols.json         # full symbol index (functions, classes, types...)
  temporal.json        # git coupling, hotspots, bug history
  hotspots.md          # risk-ranked files (static, always available)
  AGENT.md             # tool usage guide for AI agents
  modules/*.md         # per-module structural analysis
  decisions/*.md       # architectural decision records
  sessions/*.md        # session change logs
  patterns.md          # coding patterns and conventions

CLAUDE.md              # ← inline context injected here
.cursorrules           # ← and here (if exists)
.windsurfrules         # ← and here (if exists)

Six Knowledge Layers

MCP Tools (5)

Five focused tools that provide capabilities agents can't get from reading code:

MCP Resources (3)

Static knowledge available without tool calls:

MCP Prompts (2)

CLI Commands

How It Works

Hybrid extraction: tree-sitter native N-API for structure (symbols, imports, calls across 27 languages) + host LLM for semantics (what modules do, why they're built that way). Zero extra API keys.

Inline context injection: After analysis, CodeCortex injects a rich knowledge section directly into CLAUDE.md and other agent config files. This includes architecture overview, risk map with coupled file names, and editing directives — so agents have project context from the first prompt without needing MCP.

Git hooks keep knowledge fresh — codecortex update runs automatically on every commit, re-extracting changed files and updating temporal analysis.

Size-adaptive responses — CodeCortex classifies your project (micro → extra-large) and adjusts response caps accordingly. A 23-file project gets full detail. A 6,400-file project gets intelligent summaries. Every MCP tool response stays under 10K chars.

Supported Languages (27)

Tech Stack

• TypeScript ESM, Node.js 20+
• tree-sitter (native N-API) + 27 language grammar packages
• @modelcontextprotocol/sdk - MCP server
• commander - CLI
• simple-git - git integration
• yaml, zod, glob

License

MIT