# Whale

A terminal-first AI coding agent built specifically for DeepSeek, reporting roughly 98 percent prompt-cache hit rates and 1M-token context, with JavaScript-scriptable multi-agent workflows and MCP tool support.

Whale is an AI coding agent built natively around DeepSeek's long context, tool calling and cost efficiency, rather than being a generic multi-model wrapper, and reports around a 98 percent prompt cache hit rate that the project says cuts most sessions' cost substantially. It runs as an interactive terminal UI after a setup step that stores a DeepSeek API key, reading files, running commands, editing code and searching the web from a typed question.

Its Dynamic Workflows feature lets a user write JavaScript scripts that orchestrate multiple agents, fanning out parallel research tasks, running multi-perspective reviews, or building pipelines, and the project states that workflow scripts written for Claude Code work as-is in Whale. It also supports over 1,000 MCP servers for extended tooling such as file operations, shell, Git and web access, plus installable community skills and plugins for tasks like code review or Git workflows.

It installs with npm, Homebrew, a shell script, or a Windows PowerShell installer, and runs on macOS, Linux and Windows.

Website: https://whale-site.pages.dev