Whale is an AI coding agent for DeepSeek, positioned as terminal-first and optimized around DeepSeek's prefix-cache. The official introduction calls it "blazingly fast," open source, and extremely cost-efficient.
Main purpose and positioning: Whale is used in the terminal for code reading, file editing, command execution, and multi-agent workflow orchestration. It does not depend on an IDE and can work directly with the user's existing shell, git, and test commands. Users can start in the terminal and expand to desktop and beyond.
Main capabilities:
- Extreme cost efficiency: Natively optimized for DeepSeek, with a prefix-cache hit rate of about 90%. The official figures give a typical 5-turn coding session cost of about $0.001, about 93% cheaper than Claude Code.
- Terminal-native: No IDE dependency, offering both TUI and CLI modes, supporting interactive or headless use.
- Extensibility: External tools can be connected through MCP servers; workflows can be reused through local Agent Skills; coding agents can be customized without forking.
- Cache-first architecture: Its loop is designed around DeepSeek's byte-stable prefix-cache, so long sessions can continuously benefit from the cached prefix and remain low-cost.
- Dynamic workflows: Multi-agent orchestration scripts can be written in JavaScript, supporting fan-out research, multi-perspective review, pipeline processing, and more. It is officially said to be compatible with Claude Code and can run as-is.
How to use: The official three-step quick start is provided: install, configure, and start coding. Installation methods include curl (recommended) and PowerShell (recommended). After installation is complete, an API key must be configured.
Applicable scenarios: Suitable for long coding sessions, developers who need cost control, and workflows that want to be terminal-centric and combine existing shell/git/test tools for development. It is also suitable for scenarios that require multi-agent orchestration or extending capabilities through MCP/Skills.
Limitations: The current materials do not provide the specific contents of the complete installation commands, the range of supported operating systems, the specific DeepSeek model version requirements, the syntax specification for Skills, the specific features and availability status of the desktop version, or the detailed A of multi-agent orchestrationPI documentation. Some information is not yet available.
