⚡ Chakra AI

     ┌──────────────────────────────────────────────┐
     │  ▐▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▌  │
     │  ▐▓▌  ⚡  C H A K R A   A I              ▐▓▌  │
     │  ▐▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▌  │
     └──────────────────────────────────────────────┘
  Made by Abhirup Guha · Info Security Solution · insec.in

❇️ Quick Start

One command. That's it.

pip install chakra-ai
chakra

On first run, auto-downloads the model (~1 GB) and benchmarks your system.

# Full one-liner (Windows)
powershell -ExecutionPolicy Bypass -Command "irm https://raw.githubusercontent.com/fir3storm/chakra-ai/main/install.ps1 | iex"

# Full one-liner (Linux/macOS)
curl -fsSL https://raw.githubusercontent.com/fir3storm/chakra-ai/main/setup.sh | bash

⚡ Speed: ~300 tokens/sec via llama.cpp GGUF — code generation takes under 1 second. ~1 GB RAM. No GPU. Offline.

✨ Features

🎮 The Terminal

⟡ [SYSTEM] [INFO] Tuesday, August 05, 2026 00:16 | win32
⟡ [ENGINE] [INFO] llama.cpp GGUF backend (fastest)
⟡ [WORKSPACE] [INFO] project: main.py (234L), utils.py (45L)

(fullstack) > make a python calculator

  ⠋ Thinking about: make a python calculator...
  ✔ Thinking... 28 lines (0.9s)
  💾 chakra_output/generated_script.py → 28 lines
  ⚡ Sandbox execution → Exit 0

(fullstack) > /plan build a REST API with Flask
  Step 1/4: app.py with routes       ✔
  Step 2/4: database.py with models  ✔
  Step 3/4: auth.py with JWT         ✔
  Step 4/4: test_api.py              ✔

(fullstack) > /git status
  M app.py  M database.py

(fullstack) > /git commit -m "Add REST API modules"
  ✔ Committed: a1b2c3d

🔧 All Commands

🏗 Three Engine Tiers

┌──────────────────────────────────────────────────────────────┐
│ ENGINE A            ENGINE B            ENGINE C (default)   │
│ kimi-k3-in-c        PyTorch K3          llama.cpp GGUF       │
│                                                                  
│ Full 2.78T model    Full 2.78T model    Qwen2.5-Coder 1.5B   │
│ 8.24 GB RAM         8-10 GB RAM         1 GB RAM             │
│ 1.56 TB disk        1.56 TB disk        1 GB disk            │
│ Linux only          Windows/Linux       Any OS               │
│ ─────────────────   ─────────────────   ─────────────────    │
│ --trunk <path>      --trunk <path>      pip install chakra-ai│
└──────────────────────────────────────────────────────────────┘

🔒 Security Auditing

10+ OWASP rules: hardcoded credentials, SQL injection, eval()/exec(), os.system, subprocess(shell=True), MD5/SHA1, weak ciphers, unsafe deserialization, dynamic imports.

⚙️ CLI Flags

chakra [OPTIONS]
  --preset   {laptop,desktop,workstation,server}
  --engine   {auto,c-backend,pytorch,local}
  --gen      INT     Max tokens (default: 512)
  --prompt   TEXT    Single-shot mode
  --device   {cpu,cuda}

🌏 Made in India 🇮🇳

Chakra AI is built in India. "Chakra" (चक्र) means "wheel" — the continuous cycle of code generation, execution, auditing, and refinement.

Author: Abhirup Guha · Organization: Info Security Solution · Web: insec.in

📖 Architecture

Speed

• llama.cpp GGUF — Q4_K_M quantized, SIMD kernels, 300+ tok/s
• Float16 — PyTorch fallback at half precision
• Multi-threaded — All CPU cores

Kernel (Engine B — Full Kimi K3)

• Fused MXFP4 Matmul — 7.5x less memory traffic
• Ring Buffer Trunk Streaming — Pinned prefix + ring slot
• Direct I/O — O_DIRECT / FILE_FLAG_NO_BUFFERING

chakra-ai/
├── chakra/               # Core engine (19 files)
│   ├── agent.py          # Sandbox, self-debugging
│   ├── cli.py            # REPL, /plan, /edit, /git, /memory
│   ├── engine_llama.py   # llama.cpp GGUF backend
│   ├── engine_c_backend.py
│   ├── model.py          # PyTorch K3 + MXFP4Linear
│   ├── multi_agent.py    # Multi-agent orchestrator
│   ├── security.py       # InfoSecAuditor (OWASP)
│   └── ui.py             # Modern terminal UI
├── tests/                # Test suite
├── install.ps1           # Windows one-liner
├── setup.sh              # Linux one-liner
└── README.md

🙏 Acknowledgments

Built on the pioneering work of Fareed Khan and kimi-k3-in-c — a 176 KB C99 binary proving 2.78T parameter inference on 8GB RAM.

⚖ License

MIT License · © 2026 Abhirup Guha · Info Security Solution · insec.in · Made in India 🇮🇳