Whale

简体中文 · English

Whale is an unofficial DeepSeek CLI / DeepSeek coding agent for the terminal.
It can read code, edit files, run commands, and extend the agent with MCP and Skills.

90% live prefix-cache hit · ~30x cheaper per task vs Claude Code · terminal-first · open source

📋 Roadmap · View current direction and available tasks

---

Quick Start

Install with the script:

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

Install with Homebrew:

brew install usewhale/tap/whale

Windows PowerShell:

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

Windows CMD (Command Prompt):

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

The installer prefers the Windows x64 or ARM64 package for your system and adds whale.exe to the current user's PATH. If an older release does not include an ARM64 package, it falls back to x64 emulation. You can also download the matching Windows zip manually from GitHub Releases.

First run:

whale setup
whale doctor
whale

Upgrade:

brew upgrade usewhale/tap/whale
# or rerun the install script

Whale currently uses the DeepSeek API. Before running Whale, create an API key in the DeepSeek Platform. See the DeepSeek API docs for API details.

Platform support: Whale currently supports macOS, Linux, and Windows.

You can also run a one-shot prompt:

whale exec "Explain what this repository does"
printf 'Summarize the current directory\n' | whale exec

---

How It Compares

	Whale	Claude Code	Codex CLI	Cursor	Aider
Primary interface	Terminal TUI/CLI	Terminal agent	Terminal agent	IDE	CLI
Default backend	DeepSeek	Anthropic	OpenAI	Multi-model	Multi-model
DeepSeek optimized	yes	no	no	no	limited
Prefix-cache friendly	yes	n/a	n/a	model-dependent	limited
Local code read/write	yes	yes	yes	yes	yes
Shell / test execution	yes	yes	yes	partial	yes
/ask read-only mode	yes	partial	partial	n/a	partial
/plan planning mode	yes	yes	yes	n/a	partial
MCP	yes	yes	version-dependent	partial	partial
Skills / reusable workflows	yes	yes	yes	partial	limited
Open source	yes	no	yes	no	yes

Whale is not trying to support every model. Its focus is turning the DeepSeek API into a stable, low-cost local coding agent that can stay open for long development sessions.

Why DeepSeek-only?

DeepSeek's low token price is only part of the story. The real advantage for long-running coding agents is prefix caching.

DeepSeek's prefix cache is sensitive to byte stability. Whale's loop is designed around that constraint: append-only turns, stable context ordering, and recoverable session records help long tasks keep benefiting from cached prefixes.

That is why Whale is not rushing toward a generic provider abstraction. Claude, OpenAI, and DeepSeek differ in cache mechanics, tool-call behavior, and reasoning controls. A generic wrapper usually hides the DeepSeek-specific parts that matter most.

Whale includes DeepSeek-specific handling for:

Generic agent assumption	What DeepSeek can do	Whale's handling
Tool-call JSON is stable	Payloads can be malformed, escaped, or mixed into reasoning	schema-guided repair / scavenge paths
Deep tool schemas survive intact	Some nested parameters may be dropped	flatter tool parameters
Failed tools should always trigger replan	Some failures should pass through to the model	finer failure classification and recovery
User cancellation is just another tool failure	Cancellation should not continue recovery or replanning	dedicated interrupt path
Reasoning depth is prompt-only	DeepSeek exposes reasoning_effort	runtime effort control

Whale validates tool inputs against the schema first, then repairs common recoverable shape errors only on failing paths: null optional fields, stringified arrays, bare strings for array fields, markdown-autolink paths, and read_file calls that provide only offset or limit. Repair and invalid-input counts are visible in /stats.

Whale's goal is to make DeepSeek's pricing, cache behavior, and coding capability usable in a real terminal workflow.

---

What Whale Can Do

• Understand codebases: read files, search code, and summarize project structure.
• Modify code: generate patches, edit files, add tests, fix bugs, and handle local refactors.
• Run commands: execute shell commands, tests, builds, and diagnostic scripts, then bring results back into the conversation.
• Work interactively: use the local TUI, persist sessions, and resume with whale resume.
• Ask read-only questions: use /ask when you want analysis without file edits.
• Plan before execution: use /plan to review a plan before letting the agent implement it.
• Extend tools: connect external tools with MCP and reuse workflows with Skills.
• Run headlessly: use whale exec from scripts, CI, or one-shot tasks.
• 1M context window: DeepSeek V4 models automatically use 1M token context with no manual config.

Common Commands

Command	Purpose
whale	Start the interactive TUI
whale setup	Save a DeepSeek API key
whale doctor	Run health checks
whale exec "prompt"	Run one prompt non-interactively
whale --worktree [name]	Create or reuse an isolated git worktree for this interactive session
whale exec --worktree [name] "prompt"	Run one prompt inside an isolated git worktree
whale --dangerously-skip-permissions	Skip tool approval prompts for this run; suitable for external sandboxes or fully trusted repos
whale migrate-config	Migrate Whale v0.1.8-or-earlier config files to config.toml
whale resume	Open the session picker
whale resume --last	Resume the most recent session
whale resume <id>	Resume a specific session
/model	Change model, reasoning effort, and thinking
/permissions	Adjust tool approval mode
/focus	Toggle focused view to hide thinking and tool details
/ask [prompt]	Read-only question mode
/plan [prompt]	Plan first, then decide whether to execute
/review [target]	Build a code-review prompt for local changes, branches, PRs, or commits
/status	Show current session, mode, model, and config status
/compact	Compact the current conversation context
/init	Generate AGENTS.md for the current repository
/skills	Open the Skills menu to list, insert, or enable/disable local skills
/plugins	Open the built-in plugin manager to enable or disable official plugins
/memory	View or manage long-term memory saved by the official memory plugin
/mcp	Show MCP server status

MCP

Whale can load external tools from MCP servers.

See docs/mcp.md for setup and supported features.

Skills

Whale supports local Agent Skills for reusable workflows, team conventions, or tool-specific guidance.

In the TUI, type $ to search and insert a $skill-name. You can also run /skills: List skills opens the same $ picker and inserts the selected skill into the composer, while Enable/Disable Skills opens a searchable toggle manager.

See docs/skills.md for details.

Plugins / Memory

Whale currently ships with the official memory plugin for saving and recalling long-term memory. Use /memory to view or manage memories, and /plugins to open the plugin manager and press Space to enable or disable plugins.

See docs/plugins.md for details.

Configuration

Whale uses ~/.whale/config.toml for global settings, ./.whale/config.toml for shared project settings, and ./.whale/config.local.toml for private project-local overrides. Config loads in this order:

defaults < global < project shared < project local < CLI flags/env

Run this only if you used Whale v0.1.8 or earlier and have local preferences.json or settings.json files:

whale migrate-config

If you started with Whale v0.1.9 or newer, you do not need this command.

See docs/configuration.md for details.

Worktrees

In a git repository, use --worktree to create or reuse an isolated worktree for the current run:

whale --worktree feature-x
whale exec --worktree feature-x "implement and test this change"

Whale stores managed worktrees under ./.whale/worktrees/<name> and uses branches named worktree-<name>. If you omit the name, Whale generates a session-* name. On creation, Whale best-effort copies only ./.whale/config.local.toml; it does not copy session logs, API keys, private MCP config, or the whole ./.whale directory.

When you exit an interactive worktree session, Whale removes a clean worktree automatically. If the worktree has uncommitted files or commits after the original checkout head, Whale prompts you to keep or remove it. Removing a worktree discards that checkout and its uncommitted changes, but it does not delete the conversation — whale resume <id> resumes from the original workspace.

whale exec --worktree runs non-interactively and leaves its worktree on disk. To review and clean it up, re-enter it interactively with whale --worktree <name> and exit, which runs the same cleanup prompt. This version does not yet include tmux or stale sweeping.

Coding Plan Support

Yes. Whale connects to Coding Plan / third-party DeepSeek-compatible endpoints via a custom API endpoint that supports /chat/completions.

For Alibaba Cloud Bailian, configure ~/.whale/config.toml:

model = "deepseek-v4-flash"
reasoning_effort = "high"
thinking_enabled = true

[api]
base_url = "https://dashscope.aliyuncs.com/compatible-mode/v1"

Then set the corresponding API key:

DEEPSEEK_API_KEY=sk-... whale

Or run:

whale setup

Notes:

• base_url should be the root endpoint — do not include /chat/completions. Whale appends it automatically.
• The DEEPSEEK_BASE_URL environment variable overrides [api].base_url in ~/.whale/config.toml.
• Whale still prioritizes DeepSeek-native models, streaming, thinking, tool calls, and prefix-cache workflows. Full compatibility with third-party endpoints depends on server-side support.

---

Non-goals

• Not a generic multi-model wrapper. Whale is DeepSeek-only for now and prioritizes DeepSeek's cache, tool-call, and cost advantages.
• Not an IDE. Whale is terminal-first and works with your shell, git, and test commands instead of replacing IDEs like Cursor.

Project Status

Whale is moving quickly. It is best used first on personal projects, experimental repositories, or workflows where changes can be reviewed and rolled back.

Disclaimer: This project is not affiliated with DeepSeek Inc. It is an independent open-source community project.

Contributing

See CONTRIBUTING.md for cloning, local development, testing, issues, and pull requests.

Current development direction and available tasks are listed in ROADMAP.md.

Security

For security-sensitive issues, see SECURITY.md.