Kode

Kode is an open-source AI coding agent for the terminal, powered by Sarvam AI's free sarvam-m model.

Kode officially targets WSL, Linux, and macOS. Native Windows PowerShell or cmd.exe is not currently supported.

Why Kode?

• Free LLM: Powered by Sarvam AI's sarvam-m model - completely free for developers
• Open Source: Built transparently with TypeScript, no black boxes
• Indian-Made: Sarvam AI is an Indian company building world-class AI models
• Terminal-First: Clean, minimal CLI interface that stays out of your way
• Trustworthy: Always shows diffs before modifying files, git snapshots for undo

Install

npm install -g kode-ai

Or build from source:

git clone https://github.com/askpext/kode.git
cd kode
pnpm install
pnpm build

Setup

1. Get your free API key from Sarvam AI

2. Set the environment variable or configure in kode.json:

   export SARVAM_API_KEY=your-api-key

3. Run Kode:

   kode

Usage

Running Kode

Kode is an interactive terminal UI application. Run it directly in your terminal:

cd kode
node dist/cli.js

Basic Commands

Once inside Kode, you can:

• Type natural language requests to understand code, write features, or fix bugs
• Use slash commands for session management

Slash Commands

| Command | Description | |---------|-------------| | /help | Show all commands | | /new | Start a new session | | /sessions | List recent sessions | | /resume <id> | Resume a specific session | | /undo | Restore last git snapshot | | /clear | Clear the screen | | /model | Open interactive model switcher (sarvam-30b, 105b, etc.) | | /cost | Show token usage |

Tools

Kode has access to these tools via prompt-based calling:

• read_file - Read file contents (smart chunking for large files)
• write_file - Write files (always shows diff first)
• edit_file / replace_multi - Targeted and multi-block text replacement
• bash - Run shell commands synchronously
• bash_background / bash_status - Spawn and manage long-running dev servers
• grep - Search codebase with ripgrep
• fetch_url - Native web fetching (HTML to Markdown)
• list_dir - List directory contents
• todo_write / todo_read - Planning and task tracking

Configuration

Create a kode.json in your project root:

{
  "provider": {
    "apiKey": "your-sarvam-api-key",
    "baseUrl": "https://api.sarvam.ai/v1",
    "model": "sarvam-m"
  },
  "permission": {
    "bash": "ask",
    "write": "ask",
    "edit": "ask"
  },
  "context": {
    "maxTokens": 28000,
    "compressAt": 0.80
  }
}

Permission options: ask, allow, deny

Project Guidelines (AGENTS.md)

Kode automatically reads AGENTS.md or CLAUDE.md from your project root and appends it to the system prompt. Use this file to document:

• Project structure
• Coding conventions
• Architecture decisions
• Testing requirements

Key Features

1. Native Web Browsing - Can fetch documentation and issues directly from URLs
2. Background Processes - Run watch servers and test runners asynchronously
3. Interactive Model Switcher - Hot-swap between Sarvam models dynamically via TUI
4. Never silently modifies files - Always shows colored diffs before writing
5. Smart large file handling - Chunks files over 8000 tokens, never crashes
6. Git snapshots - Takes a snapshot before every write for instant undo
7. Context compression - Automatically summarizes at 80% of 32k token limit

Tech Stack

• Runtime: Node.js + TypeScript
• Terminal UI: Ink v5 (React for CLIs)
• LLM: Sarvam AI (sarvam-m model) - uses prompt-based tool calling
• Storage: sql.js (SQLite in pure JavaScript)
• Shell: execa
• Config: cosmiconfig

Documentation

• Slash Commands Reference
• Architecture Overview
• Configuration Guide
• Agent Behavior Guide
• Contributing Guidelines
• Usage Examples

License

MIT License - See LICENSE for details.

Get Your API Key

Visit https://sarvam.ai to get your free API key.

Built with ❤️ for developers