⚡ Chakra AI

  ╔═══════════════════════════════════════════════════════════════╗
  ║   ____ _   _    _    _  ___  _       _    ___                ║
  ║  / ___| | | |  / \  | |/ / |  _ \   / \  |_ _|  CHAKRA AI   ║
  ║ | |   | |_| | / _ \ | ' /| |_) | / _ \  | |   Agentic Code  ║
  ║ | |___|  _  |/ ___ \| . \|  _ < / ___ \ | |   Terminal       ║
  ║  \____|_| |_/_/   \_\_|\_\_|_| \_\_/   \_\___|               ║
  ╚═══════════════════════════════════════════════════════════════╝

🎯 What is Chakra AI?

Chakra AI is a terminal-based AI coding companion built from scratch in India. Type what you want in plain English, and an AI agent writes code, executes it in an isolated sandbox, audits it for security vulnerabilities, and self-debugs if anything breaks — all 100% offline on an 8GB RAM laptop.

You: "make a folder called Abhi and put a calculator in it"

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

Done. Folder 'Abhi' created with calculator.py inside.

No API keys. No internet. No GPU. Just your terminal, an AI, and ~1 GB RAM.

❇️ Quick Start

One command. That's it.

# Windows PowerShell (Run as Administrator)
powershell -ExecutionPolicy Bypass -Command "irm https://raw.githubusercontent.com/fir3storm/chakra-ai/main/install.ps1 | iex"

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

After install, just type:

chakra

The installer handles everything — Python deps, C++ compilation, model download (~1 GB), system benchmark. Takes ~5 minutes.

⚡ Speed: On a standard laptop CPU, Chakra AI generates ~300 tokens/sec via llama.cpp with a GGUF-quantized model — code generation takes under 1 second.

✨ Features

🎮 The Terminal Experience

(fullstack) > hi

  ⠋ Thinking...
  Hello! I'm your coding assistant. What would you like to build today?
  ✔ Thinking... 3 chunks (0.3s)

(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) > /persona infosec
(fullstack) > /audit chakra_output/generated_script.py

  🛡 InfoSec Audit Report
  Target: chakra_output/generated_script.py
  ✔ PASS: Score: 100/100

(fullstack) > /status

  Engine:   llama.cpp (Qwen2.5-Coder-1.5B Q4_K_M)
  Threads:  22
  Persona:  [INFOSEC] - InfoSec Expert
  RAM:      ~1 GB

🏗 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               │
│ Bit-exact verified  atol 1e-4           300+ tok/s           │
│ ─────────────────   ─────────────────   ─────────────────    │
│ --trunk <path>      --trunk <path>      setup.bat / setup.sh │
└──────────────────────────────────────────────────────────────┘

Engine C ships with a setup.bat/setup.sh one-click installer.
Engine A runs the legendary 176 KB kimi-k3-in-c binary as a subprocess.
Engine B is the PyTorch implementation for Windows when you have the full checkpoint.

🔧 All Commands

🔒 Security Auditing

Chakra AI scans every generated code file with 10+ OWASP rules:

• 🔑 Hardcoded credentials — passwords, API keys, tokens
• 💉 SQL injection — dynamic query construction
• ⚠️ Code injection — eval(), exec(), __import__
• 🔨 Command injection — os.system, subprocess(shell=True)
• 🔐 Weak cryptography — MD5, SHA1, DES
• 📦 Unsafe deserialization — pickle.loads, yaml.unsafe_load

Each finding gets a severity rating, line number, and actionable remediation.

⚙️ CLI Flags

python -m chakra.cli [OPTIONS]

  --preset   {laptop,desktop,workstation,server}  Hardware preset
  --engine   {auto,c-backend,pytorch,local}        Inference engine
  --trunk    PATH                                  Trunk weights (Engine A/B)
  --trunk-gb FLOAT                                 Memory budget for streaming
  --gen      INT                                   Max tokens (default: 512)
  --prompt   TEXT                                  Single-shot mode (no REPL)
  --tiny                                          13-layer test model
  --device   {cpu,cuda}                            Compute device

🌏 Made in India 🇮🇳

Chakra AI is proudly built in India. The name "Chakra" (चक्र) is a Sanskrit word meaning "wheel" or "cycle" — representing the continuous cycle of code generation, execution, auditing, and refinement.

• Author: Abhirup Guha
• Organization: Info Security Solution
• Location: India

Why "Chakra"?

In Indian philosophy, chakras are energy centers. Each Chakra AI component maps to one:

📖 Technical Architecture

Speed Optimizations

• llama.cpp GGUF backend — Q4_K_M quantized model, SIMD-optimized C++ kernels, 300+ tok/s
• Float16 precision — PyTorch fallback at half precision, 3 GB RAM instead of 6 GB
• Multi-threaded — Uses all CPU cores for matrix operations
• torch.inference_mode() — Faster than no_grad() for generation

Kernel-Level (Engine B — Full Kimi K3)

• Fused MXFP4 Matmul — Packed 4-bit nibbles, 7.5x less memory traffic
• Ring Buffer Trunk Streaming — Pinned prefix + ring slot for 1.56 TB checkpoint
• Direct I/O Reader — O_DIRECT (Linux), FILE_FLAG_NO_BUFFERING (Windows)
• Bit-Exact Verification — Gate ladder tests at atol=1e-4

Repository Structure (48 files)

chakra-ai/
├── chakra/                     # Core engine (18 files)
│   ├── agent.py                # Sandbox, self-debugging
│   ├── cli.py                  # REPL terminal & commands
│   ├── engine_c_backend.py     # kimi-k3-in-c wrapper
│   ├── engine_llama.py         # llama.cpp GGUF backend ⚡
│   ├── model.py                # PyTorch K3 + MXFP4Linear
│   ├── multi_agent.py          # Multi-agent orchestrator
│   ├── ops.py                  # Fused kernels, SiTU-GLU, KDA
│   ├── security.py             # InfoSecAuditor (OWASP)
│   ├── session.py              # Session persistence
│   ├── persona.py              # Persona management
│   ├── trunk_streamer.py       # Ring buffer streaming
│   ├── st_reader.py            # Safetensors + DirectReader
│   └── ui.py                   # Spinner, tools, chat roles
├── tests/                      # Test suite
├── tools/                      # Benchmark + model downloader
├── setup.bat / setup.sh        # One-click installer
├── start_chakra_ai.bat         # Windows launcher
└── README.md

🙏 Standing on the Shoulders of Giants

Chakra AI builds upon the pioneering work of Fareed Khan and his remarkable kimi-k3-in-c project — a 176 KB C99 binary proving a 2.78 trillion parameter model runs on 8GB RAM. His innovations in zero-copy streaming, fused MXFP4 matmul, O_DIRECT I/O, and bit-exact verification set the gold standard.

Chakra AI extends that vision with a complete agentic coding system while remaining honest about where it differs. To Fareed Khan: thank you. 🙏

⚖ License

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