🐙 Kraken Code

An AI coding agent for your terminal — chat, generate, edit, explain, and refactor code in JavaScript, TypeScript, and Python.

Kraken Code is a modern command-line agent that connects to the Kraken AI proxy and gives you a beautiful, minimal interface to over 30 frontier models — without ever asking for an API key.

✨ Features

• 💬 Interactive chat in your terminal with streaming responses
• 🛠️ Rich tool set — read, write, patch, search, grep, move, delete, fetch URLs, run commands
• 🧠 30+ models across OpenAI, Anthropic, Google, DeepSeek, xAI, Qwen, Kimi, Llama
• 🎨 Beautiful UI — gradients, spinners, syntax highlighting, markdown rendering
• 🔌 No API key required — points to the Kraken AI proxy out of the box
• 📁 Project-aware — gitignore-aware file listing, attach files to context
• ⚡ Tool calling with automatic fallback for proxies that don't support it
• 🧩 Clean architecture — separated commands, services, client, UI

📦 Installation

From source

git clone <your-repo-url> kraken-code
cd kraken-code
npm install
npm run build
npm link

Now kraken is available globally:

kraken --help

Local development

npm install
npm run dev -- chat

🚀 Quick start

# Start interactive chat
kraken

# Ask one question and exit
kraken ask "How do I write a TypeScript generic that constrains to objects?"

# Generate a new file
kraken generate src/utils/slugify.ts -p "create a slugify function with diacritics removal"

# Edit an existing file
kraken edit src/server.ts -p "add request logging middleware"

# Explain code
kraken explain src/core/agent.ts

# Get a code review
kraken review src/api/users.py

# Refactor a file
kraken refactor src/parser.js -p "convert to TypeScript with strict types"

# Browse and select models
kraken models
kraken models set

🧭 Commands

| Command | Description | |---|---| | kraken / kraken chat | Start the interactive chat REPL | | kraken ask <question> | One-shot question, prints answer and exits | | kraken generate <file> -p "..." | Generate a new file from a prompt | | kraken edit <file> -p "..." | Edit a file according to an instruction | | kraken explain <file> | Explain what a file does | | kraken review <file> | Get review suggestions for a file | | kraken refactor <file> -p "..." | Refactor a file with an instruction | | kraken models | List all available models | | kraken models set | Interactively pick the default model | | kraken config show | Show current configuration | | kraken config get <key> | Get a single config value | | kraken config set <key> <value> | Set a config value | | kraken config reset | Reset config to defaults | | kraken config path | Print the config file path |

Global flags

| Flag | Description | |---|---| | -m, --model <name> | Override model for this invocation | | -c, --cwd <path> | Working directory (defaults to current) | | --no-stream | Disable streaming | | --debug | Enable debug output | | -v, --version | Show version | | --help | Show help |

Slash commands inside chat

| Slash | Description | |---|---| | /help | Show available slash commands | | /model | Switch model mid-conversation | | /clear | Reset conversation history | | /files | List project files | | /add <file> | Attach a file to your next message | | /save <path> | Save the last response to a file | | /exit or /quit | Leave Kraken |

🤖 Available models

Kraken Code ships with the full Kraken AI gateway model catalogue:

OpenAI — gateway-gpt-5, gateway-gpt-5-1, gateway-gpt-5-3, gateway-gpt-5-4, gateway-gpt-5-5, gateway-gpt-5-mini, gateway-gpt-5-nano, gateway-gpt-5-online, gateway-gpt-4o, gateway-gpt-4-1-mini, gateway-gpt-4-1-nano, gateway-o3, gateway-o3-mini, gateway-o4-mini

Anthropic — gateway-claude-opus-4-7 (default), gateway-claude-opus-4-6, gateway-claude-opus-4-5, gateway-claude-opus-4-1, gateway-claude-sonnet-4, gateway-claude-sonnet-4-6

Google — gateway-gemini-3-1-pro, gateway-gemini-3-pro, gateway-gemini-2-5-pro, gateway-gemini-2-5-flash

DeepSeek — gateway-deepseek-v4-pro, gateway-deepseek-v4-flash, gateway-deepseek-r1, gateway-deepseek-v3

xAI — gateway-grok-4, gateway-grok-3

Alibaba — gateway-qwen-3-max, gateway-qwen-qwq-32b

Moonshot — gateway-kimi-k2

Meta — gateway-llama-3-3-70b

⚙️ Configuration

Configuration lives at the platform-appropriate user config directory (run kraken config path to see it).

Default settings

| Key | Default | Description | |---|---|---| | endpoint | https://krakenai.onrender.com/v1 | Proxy base URL | | defaultModel | gateway-claude-opus-4-7 | Default model | | stream | true | Stream responses | | autoApproveWrites | false | Skip write confirmations | | autoApproveShell | false | Allow run_command tool to auto-execute | | maxContextFiles | 20 | Max files in project context | | maxContextBytes | 200000 | Max total context bytes | | timeoutMs | 120000 | HTTP timeout | | retries | 3 | Retry attempts on network/5xx errors | | theme | ocean | Color theme |

Examples

# Change the default model
kraken config set defaultModel gateway-gpt-5

# Point at a different proxy (e.g., self-hosted)
kraken config set endpoint https://my-proxy.example.com/v1

# Enable auto-approve for file writes (use with care)
kraken config set autoApproveWrites true

# Reset everything
kraken config reset

💡 Example workflows

Chat with a file attached

$ kraken
you › /add src/server.ts
✓ Attached src/server.ts (1842 chars)
you › Add error handling to the /users endpoint

Kraken will read the attached file, propose changes, and either render them as Markdown or call the write_file tool to update the file (with a diff + confirmation).

Generate a Python module

kraken generate scripts/csv_to_parquet.py \
  -p "convert a CSV file to a Parquet file using pyarrow, with CLI args via argparse"

Refactor JavaScript to TypeScript

kraken refactor src/legacy.js \
  -p "rewrite in strict TypeScript with proper types and ESM imports"

Code review

kraken review src/api/users.py

🧰 Built-in tools

The agent has access to the following tools. Each tool call is rendered in a colored box in the terminal so you can see exactly what the model is doing.

| Tool | Description | |---|---| | read_file(path) | Read a single file | | read_many_files(paths) | Read multiple files in one call | | list_files(directory?) | List project files (gitignore-aware) | | write_file(path, content) | Create/overwrite a file with diff + confirm | | apply_patch(path, search, replace) | Surgical search/replace edit | | search_files(pattern, glob?) | Regex search across files | | grep(query, glob?) | Find files containing a literal string | | create_directory(path) | Create a directory | | delete_file(path) | Delete a file (asks confirmation) | | move_file(from, to) | Move or rename a file | | fetch_url(url) | HTTP GET a URL | | run_command(cmd) | Run a shell command (off by default) |

🏗️ Architecture

src/
├── index.ts            # entry point
├── cli.ts              # Commander setup
├── commands/           # one file per command
├── core/
│   ├── agent.ts        # message + tool loop
│   ├── client.ts       # HTTP client (stream + non-stream)
│   ├── tools.ts        # tool definitions and handlers
│   ├── prompts.ts      # system + task prompts
│   └── context.ts      # project context building
├── config/             # defaults, model catalogue, persistent store
├── ui/                 # theme, banner, logger, spinner, markdown, diff, box
├── fs/                 # safe file reader, writer (diff + confirm), project walker
├── types/              # API + config types
└── utils/              # errors, language detection, token estimation

The Agent maintains the message history and runs a loop:

1. Send messages (+ tool definitions) to the proxy
2. If the model returns native tool_calls, execute them and feed results back
3. If the proxy doesn't support tool calling, fall back to parsing JSON action blocks from the assistant text
4. Stop when the model produces a final text reply

🛡️ Safety

• No API key required — auth is handled by the proxy
• Write confirmations — every file write shows a unified diff and prompts before applying
• Path safety — all reads/writes are constrained to the working directory
• Shell off by default — run_command requires explicit per-invocation confirmation unless autoApproveShell is enabled

🛣️ Roadmap

• Multi-file refactors with patch planning
• Project memory (persistent context across sessions)
• Plugins / custom tools
• Cost & token telemetry
• Workspace-level .kraken.json config

📄 License

MIT © Kraken AI