CodeSage ⚡

Local AI code reviewer. No cloud. No API keys. Your code never leaves your machine.

What it does

Paste any GitHub repo URL → CodeSage reviews it like a senior engineer. Finds bugs, security issues, and performance problems with exact file names and line numbers. Everything runs on your Mac. Zero internet required after setup.

Why CodeSage

	Cursor / Copilot / GitHub Actions	CodeSage
Runs 100% locally	❌	✅
Code stays private	❌	✅
No API key needed	❌	✅
Works offline	❌	✅
Free forever	❌	✅

How it works

GitHub URL → Ingestion → Chunking → Rule Engine → Fast Pass (Qwen 1.5B) → Deep Pass (DeepSeek 6.7B) → Structured Report

1. Ingestion — clones repo, parses all code files
2. Rule Engine — deterministic checks for SQL injection, hardcoded secrets, bare excepts
3. Fast Pass — Qwen2.5-Coder-1.5B scans every file for issues
4. Deep Pass — DeepSeek-Coder-6.7B-4bit analyzes files with issues in detail
5. Memory — stores patterns across reviews, gets smarter over time

Setup

git clone https://github.com/sameersharma06/CodeSage.git
cd CodeSage
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
python3 -m api.main

Open http://localhost:8000

Models download automatically on first run (~4GB total).

Requirements

• macOS (Apple Silicon recommended)
• Python 3.10+
• 8GB RAM minimum, 16GB recommended
• Git installed

Tech Stack

• Models — Qwen2.5-Coder-1.5B (fast pass) + DeepSeek-Coder-6.7B-4bit (deep pass)
• Inference — MLX (Apple Silicon optimized)
• Backend — FastAPI
• Rule Engine — AST + regex (deterministic, never hallucinates)
• Memory — local JSON store, tracks patterns across reviews
• Frontend — vanilla HTML/CSS/JS

Architecture

codesage/ ├── core/ │ ├── ingestion.py # clone + parse repos │ ├── chunker.py # smart code splitting │ ├── agent.py # multi-pass review pipeline │ ├── model_manager.py # lazy load / unload models │ ├── rule_engine.py # deterministic bug detection │ └── memory.py # pattern memory across reviews ├── api/ │ └── main.py # FastAPI server └── ui/ ├── index.html ├── style.css └── app.js

Performance

• Reviewed 80 file repo in ~3 minutes on M-series Mac
• Peak RAM: ~7GB (well within 16GB)
• No GPU cloud costs

Roadmap

• GitHub Actions integration
• VS Code extension
• Multi-language support (Go, Rust, Java)
• Team workspace (self-hosted)
• PR review mode

Built by

Sameer Sharma — First year CS & AI student, Haryana India Building local AI systems on Apple Silicon.

GitHub · LinkedIn