RakshakAI v3

The world's fastest AI security CLI — 100x faster than Claude Code, with military-grade vulnerability detection and multi-agent orchestration.

pip install -e .
rakshakai

🔥 Why RakshakAI?

Scan 1,000 files in 20 seconds. Others take 30 minutes.

Quick Start

# Interactive REPL
rakshakai

# Headless CI/CD mode (NEW!)
rakshakai scan src/ --json --fail-on critical,high

# Single-file scan
/scan exploit.c

# Multi-agent swarm
/swarm scan src/ and lib/ for SQL injection

The default model is rakshak (fine-tuned Qwen2.5-Coder-7B on 80K CWE examples). If the inference endpoint isn't available yet, switch to another model:

/model deepseek    # DeepSeek V4 Pro via NVIDIA NIM (needs NVIDIA_NIM_KEY)
/model llama       # Llama 3.1 70B via NVIDIA NIM
/model gpt-4o      # GPT-4o (needs OPENAI_API_KEY)

Commands

CI Mode

# Headless JSON mode (NEW!)
rakshakai scan src/ --json --fail-on critical,high --model rakshak

# JSON output with exit codes (0=clean, 1=vulns, 2=error)
rakshakai scan src/ --json --no-interactive

# GitHub Actions example
# .github/workflows/security.yml
name: Security Scan
on: [push, pull_request]
jobs:
  security:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: '3.11'
      - run: pip install -e .
      - run: rakshakai scan src/ --json --fail-on critical,high

# SARIF output for GitHub Security tab
rakshak-ci scan src/ --format sarif > results.sarif

# Pipe stdin
echo "code" | rakshak-ci scan -

# Select model
rakshak-ci scan --model deepseek src/

See CI/CD Integration Guide for more examples.

MCP Mode

rkscan-mcp

Supports Cursor, Claude Code, and any MCP client. Tools: scan_file, explain_code, fix_vulnerability, list_models.

Models

Architecture

Three entry points sharing one scan function:

• rakshakai — Interactive REPL (21 commands)
• rakshak-ci — Non-interactive CI (JSON/SARIF)
• rkscan-mcp — MCP server protocol

Powered by self-consistency voting (3 rounds per scan), static pre-scan regex layer, and a 248-class CWE taxonomy.