bhai-cli 0.2.1

BHAI — The Dual-Brain AI Coding Agent with Punjabi Swagger. Indic-native voice + universal LLM tool-calling.

🇮🇳 BHAI-CLI

The Dual-Brain AI Coding Agent — 22 Indian Languages. One CLI.

भाई · ভাই · ભાઈ · ਭਾਈ · భాయ్ · ಭಾಯ್ · भाऊ · ভাই

BHAI is a lean, zero-state resilient CLI tool that acts as your 10x Senior SDE from India. It uses a Dual-Brain Architecture — a VibeLLM for persona and intent parsing, and a CoderAgent for tool-calling execution — all powered by litellm for universal LLM provider support and Sarvam AI for 22 Indian language understanding.

"You are BHAI, a 10x Senior SDE from India. Sab bhasha samajhte ho. You roast O(N²) loops and unoptimized Dockerfiles."

---

✨ Features

• 🧠 Dual-Brain Architecture: VibeLLM (persona + intent) → CoderAgent (tool execution) → VibeLLM (swagger summary)
• 🇮🇳 22 Indian Languages: Speak in Hindi, Punjabi, Bengali, Tamil, Telugu, Gujarati, Marathi, Kannada, Malayalam, and more — powered by Sarvam AI
• 🌍 Universal LLM Support: Every provider on Earth — OpenAI, Anthropic, Groq, Google Gemini, Sarvam, Mistral, Cerebras, Ollama, Bedrock, and more via litellm
• 🎙️ Indic Voice Interface: Sarvam Saaras v3 (STT) translates any Indian language to English, Bulbul v3 (TTS) reads responses back
• 🔧 MCP-Style Tools: execute_bash, web_search, codebase_context — with a hard [Y/n] guardrail on every command
• ⚡ Zero-State Resilience: Boots in <300ms, persistent YAML config, hot-swappable model overrides
• 📦 PyPI-Ready: Clean src/ layout, Poetry-managed, lean dependency footprint

---

🚀 Installation

pip install bhai-cli

With voice support (optional):

pip install "bhai-cli[audio]"
# Also needs: sudo apt install libportaudio2  (Linux)

---

⚙️ Setup

bhai-code setup

The interactive wizard will configure:

1. VibeLLM — Your persona brain (e.g., sarvam/sarvam-105b, groq/llama-3.3-70b-versatile)
2. CoderAgent — Your tool-calling brain (e.g., openai/gpt-4o, anthropic/claude-sonnet-4-20250514)
3. Sarvam AI — Voice interface API key (optional)

Model Format

Uses litellm's model naming:

Provider	Format	Example
OpenAI	openai/model	openai/gpt-4o
Anthropic	anthropic/model	anthropic/claude-sonnet-4-20250514
Google	gemini/model	gemini/gemini-2.5-flash
Groq	groq/model	groq/llama-3.3-70b-versatile
Sarvam	sarvam/model	sarvam/sarvam-105b
Mistral	mistral/model	mistral/mistral-large-latest
Cerebras	cerebras/model	cerebras/llama-3.3-70b
Ollama	ollama/model	ollama/llama3

---

💬 Usage

Text Mode

# Single prompt
bhai-code text "find all TODO comments in this project"

# Interactive chat
bhai-code text "what does this codebase do?" --interactive

# Override models on the fly
bhai-code text "optimize this Dockerfile" --vibe-model groq/llama-3.3-70b-versatile --coder-model openai/gpt-4o

Voice Mode

# Live microphone
bhai-code listen

# From audio file
bhai-code listen --file recording.wav

View Config

bhai-code config

---

🏗️ Architecture

User Input (Any of 22 Indian Languages / English)
         │
         ▼
┌─────────────────────┐
│   VibeLLM (Brain A)  │  ← Persona + Intent Parsing
│   Sarvam-105B / any  │
└──────────┬──────────┘
           │ {"task": "clean English intent"}
           ▼
┌─────────────────────┐
│  CoderAgent (Brain B)│  ← Tool-Calling Execution
│  Any litellm model   │
│                      │
│  Tools:              │
│  • execute_bash      │  ← [Y/n] guardrail
│  • web_search        │  ← DuckDuckGo
│  • codebase_context  │  ← AST signatures
└──────────┬──────────┘
           │ Results
           ▼
┌─────────────────────┐
│   VibeLLM (Brain A)  │  ← Swagger Summary
└─────────────────────┘

---

🔧 Tools

Tool	Description
execute_bash	Run shell commands with mandatory [Y/n] confirmation, timeout handling
web_search	DuckDuckGo search for real-time debugging and docs lookup
codebase_context	Smart directory scanner with AST-based signature extraction (anti-context-rot)

---

📁 Project Structure

bhai-cli/
├── src/bhai_cli/
│   ├── __init__.py          # Package metadata
│   ├── cli.py               # Typer CLI (setup, text, listen, config)
│   ├── config_manager.py    # Pydantic YAML/env config engine
│   ├── orchestrator.py      # Dual-Brain routing logic
│   ├── audio_engine.py      # Sarvam STT/TTS + mic recording
│   └── tools/
│       ├── __init__.py      # Tool registry + JSON schemas
│       ├── bash.py          # execute_bash with guardrails
│       ├── search.py        # web_search via DuckDuckGo
│       └── codebase.py      # codebase_context with AST
├── tests/
├── pyproject.toml           # Poetry config (lean deps)
├── README.md
└── LICENSE                  # Apache-2.0

---

🔑 Environment Variables

Variable	Description
SARVAM_API_KEY	Sarvam AI API key
VIBE_API_KEY	VibeLLM provider API key
CODER_API_KEY	CoderAgent provider API key
GROQ_API_KEY	Groq API key (auto-detected by litellm)
OPENAI_API_KEY	OpenAI API key (auto-detected by litellm)
ANTHROPIC_API_KEY	Anthropic API key (auto-detected by litellm)
GEMINI_API_KEY	Google Gemini API key (auto-detected by litellm)

You can also use a .env file in your working directory.

---

🗺️ Roadmap (v2)

• MCP Server connectivity for external tool integration
• Streaming responses
• Multi-turn CoderAgent memory
• Plugin system for custom tools

---

📄 License

Apache-2.0 — see LICENSE.

Built with ❤️ from India by Madhav Kapila.