DeepSeek-Code-Whale: DeepSeek-Native AI Coding Agent

TL;DR

TL;DR: Whale is an MIT-licensed, Go-based terminal AI coding agent built specifically for DeepSeek, with ~98% prompt cache hit rate, 1M-token context, MCP support, and Claude Code-compatible dynamic workflows.

Source and Accuracy Notes

• Project page: whale-site.pages.dev
• Source repository: github.com/usewhale/DeepSeek-Code-Whale
• License: MIT (verified via GitHub API license.spdx_id)
• Language: Go
• Source last checked: 2026-06-16

What Is Whale?

Whale is a DeepSeek-native AI coding agent. Where most coding agents are multi-model wrappers, Whale is purpose-built for DeepSeek’s 1M-token context, tool calling, and pricing. The headline numbers from the README are ~98% prompt cache hit rate, a 1M-token context window, and zero bloat. The cache hit rate is the design driver: cached prompts are dramatically cheaper than uncached ones, so a coding agent that reuses context aggressively is meaningfully cheaper per session.

The agent ships three interfaces: a TUI for interactive sessions, a CLI for one-shot questions and code reviews, and a headless mode for CI/CD, automated PR reviews, and scheduled tasks. It is not affiliated with DeepSeek Inc. — it is an independent open-source community project.

Repo-Specific Setup Workflow

Step 1: Install

Any platform via npm:

npm install -g @usewhale/whale

macOS via Homebrew:

brew install usewhale/tap/whale

Linux via the install script:

curl -fsSL https://raw.githubusercontent.com/usewhale/DeepSeek-Code-Whale/main/scripts/install.sh | sh

Windows PowerShell (Windows 10 / Windows Server 2016 or later):

irm https://raw.githubusercontent.com/usewhale/DeepSeek-Code-Whale/main/scripts/install.ps1 | iex

Windows CMD:

powershell -NoProfile -ExecutionPolicy Bypass -Command "irm https://raw.githubusercontent.com/usewhale/DeepSeek-Code-Whale/main/scripts/install.ps1 | iex"

Step 2: Configure

whale setup

This stores your DeepSeek API key and writes .whale/config.local.toml. See docs/configuration.en.md if you need a different model provider, a proxy, or a custom config.

Step 3: Launch

whale                # interactive TUI
whale ask "..."      # one-shot CLI
whale --headless     # CI/CD and automation

Deeper Analysis

Dynamic workflows

The most distinctive feature is Dynamic Workflows. A workflow is a JavaScript file under .whale/workflows/ that orchestrates multiple agents. The example in the README is a fan-out research flow that runs two parallel research agents and then synthesizes the results through a third agent. The same script can do multi-perspective review, pipeline processing, or adversarial validation. Workflows are disabled by default; enable them with /config in the TUI or by adding [workflows] enabled = true to .whale/config.local.toml.

The same workflows are designed to be Claude Code compatible, so a workflow written for Claude Code works as-is in Whale. That makes Whale a useful drop-in for teams that want to keep a single workflow format across both engines.

MCP, skills, plugins, hooks, subagents

Whale exposes the full extension surface. MCP servers let you plug in 1,000+ tools through the MCP ecosystem — databases, APIs, browser automation, and so on. Skills load domain expertise, with code review and git worktree examples in the docs. Subagents define focused child roles such as reviewers or researchers. Plugins extend the runtime. Hooks run scripts on lifecycle events. The combination makes Whale a credible platform, not just a CLI.

The interfaces

The TUI is the canonical interface for interactive coding. The CLI is for one-shot questions, quick code reviews, and single commands. The headless mode is for CI/CD, automated PR reviews, and scheduled tasks. The same engine backs all three, so a workflow that works in the TUI works in CI without rewriting it.

Non-goals

The README is honest about the non-goals. Whale is not a multi-model shell. It is DeepSeek-first and optimized for DeepSeek’s caching, tools, and pricing. It is not an IDE. It meets you where you code, in the terminal, in a desktop wrapper, or in CI.

Practical Evaluation Checklist

• [ ] Do you have a DeepSeek API key and want a DeepSeek-native coding agent?
• [ ] Is the ~98% prompt cache hit rate and 1M-token context a meaningful win for your workflow?
• [ ] Do you want to run the same JavaScript workflows on Claude Code and Whale?
• [ ] Will you connect Whale to MCP servers for databases, APIs, or browser automation?
• [ ] Do you need a TUI, a CLI, and a headless mode backed by the same engine?
• [ ] Is the “DeepSeek-only, not a multi-model wrapper” trade-off acceptable?
• [ ] Will you keep your work in reviewable repositories where changes can be rolled back?

Security Notes

Whale is a terminal agent, so it has the same trust model as the rest of that category: the agent can read, edit, and execute code on your machine. Run it inside a project you are willing to roll back. The README is explicit that the project is “best suited for personal projects, experimental repositories, and workflows where changes can be reviewed and rolled back.”

The whale setup command writes your DeepSeek API key to .whale/config.local.toml. Do not commit that file. The MCP, skill, plugin, and hook surfaces all execute local code, so treat anything you install from those surfaces the way you would treat a third-party npm dependency.

FAQ

Q: Is Whale a multi-model wrapper? A: No. The README’s non-goals section is explicit. Whale is DeepSeek-first, optimized for DeepSeek’s caching, tools, and pricing. A “different model provider, proxy, or custom config” path is documented, but the design is single-model.

Q: How is ~98% prompt cache hit rate achieved? A: Whale reuses cached context aggressively. DeepSeek pricing for cached prompts is dramatically lower than uncached prompts, so the high cache hit rate is the cost optimization that the design is built around.

Q: Are Whale workflows compatible with Claude Code? A: Yes. Workflows written for Claude Code work as-is in Whale. Dynamic Workflows are disabled by default and have to be enabled in .whale/config.local.toml or through the /config TUI command.

Q: Does Whale replace an IDE? A: No. The README is explicit that Whale is an agent, not an IDE. It meets you in the terminal, a desktop wrapper, or CI.

Q: Is Whale affiliated with DeepSeek? A: No. The README states that the project is not affiliated with DeepSeek Inc. and is an independent open-source community project.

Conclusion

Whale is a focused bet: a DeepSeek-native coding agent written in Go, optimized for the cache hit rate that makes DeepSeek’s pricing interesting, with a JavaScript workflow engine and the MCP surface that serious agent users want. If you are already a DeepSeek user and you want a terminal agent that respects that bet, Whale is worth a serious look. If you want a multi-model wrapper or an IDE replacement, this is not the project.

Related reading: GitHub Trending tools, Developer tools, Rowboat, boo, FanBox.