An open-source AI coding assistant CLI powered by DeepSeek-V3 - Build your own "Claude Code" with DeepSeek's API.
开源 AI 编程助手命令行工具，由 DeepSeek-V3 驱动 - 使用 DeepSeek API 打造属于你自己的 "Claude Code"。
____ ____ _
| _ \ ___ ___ _ __/ ___| ___ ___| | __
| | | |/ _ \/ _ \ '_ \___ \ / _ \/ _ \ |/ /
| |_| | __/ __/ |_) |__) | __/ __/ <
|____/ \___|\___| .__/____/ \___|\___|_|\_\
____ |_| _
/ ___|___ __| | ___
| | / _ \ / _` |/ _ \
| |__| (_) | (_| | __/
\____\___/ \__,_|\___|
🇨🇳 中文
- 性价比高: DeepSeek-V3 提供 GPT-4 级别的性能，成本仅为其一小部分
- 开放架构: 完全了解 AI 编程代理的工作原理，没有黑盒
- 可扩展: 添加自定义工具、修改代理循环、自由定制一切
- 注重隐私: 代码留在本地，只有提示词发送到 API
| 功能 | 描述 |
|---|---|
| 交互式 REPL | 实时与 AI 讨论你的代码 |
| 文件操作 | 读取、写入、精准编辑文件 |
| Shell 命令 | 运行 bash 命令，带安全检查 |
| 代码搜索 | 用 glob 查找文件，用 grep 搜索内容 |
| 项目上下文 | 自动加载 DEEPSEEK.md 获取项目知识 |
| 权限系统 | 危险操作前请求许可 |
| YOLO 模式 | 使用 --yolo 跳过所有提示（类似 Claude Code） |
| 对话历史 | 跨会话保存 |
🇺🇸 English
- Cost-effective: DeepSeek-V3 offers GPT-4 level performance at a fraction of the cost
- Open architecture: Understand exactly how AI coding agents work - no black box
- Extensible: Add your own tools, modify the agent loop, customize everything
- Privacy-focused: Your code stays on your machine, only prompts go to the API
| Feature | Description |
|---|---|
| Interactive REPL | Chat with AI about your code in real-time |
| File Operations | Read, write, and surgically edit files |
| Shell Commands | Run bash commands with safety checks |
| Code Search | Find files with glob, search content with grep |
| Project Context | Auto-loads DEEPSEEK.md for project-specific knowledge |
| Permission System | Asks before dangerous operations |
| YOLO Mode | Skip all prompts with --yolo (like Claude Code) |
| Conversation History | Persists across sessions |
# Clone and install in one go
git clone https://github.com/yksanjo/deepseek-code.git
cd deepseek-code
pip install -e .pip install git+https://github.com/yksanjo/deepseek-code.git- Get your API key from DeepSeek Platform
- Add credits to your account (DeepSeek-V3 is very affordable!)
- Set the environment variable:
# Add to your shell profile (~/.bashrc, ~/.zshrc, etc.)
export DEEPSEEK_API_KEY=your_key_here
# Or create a .env file in your project
echo "DEEPSEEK_API_KEY=your_key_here" > .env# Start interactive mode
deepseek-code
# Or use the short alias
dscYou should see:
____ ____ _
| _ \ ___ ___ _ __/ ___| ___ ___| | __
| | | |/ _ \/ _ \ '_ \___ \ / _ \/ _ \ |/ /
| |_| | __/ __/ |_) |__) | __/ __/ <
|____/ \___|\___| .__/____/ \___|\___|_|\_\
____ |_| _
/ ___|___ __| | ___
| | / _ \ / _` |/ _ \
| |__| (_) | (_| | __/
\____\___/ \__,_|\___|
v0.1.0
Directory: /your/project
Type help for commands, quit to exit
──────────────────────────────────────────────────
>
deepseek-codeDeepSeek Code v0.1.0
Working directory: /path/to/your/project
> Fix the type error in utils.py
> Reading utils.py...
> edit_file: utils.py
old: "def process(data):"
new: "def process(data: dict) -> list:"
Permission required: Edit file: utils.py
Allow? [y/n/always]: y
Fixed type annotation in utils.py
> quit
deepseek-code run "Add docstrings to all functions in main.py"deepseek-code # Start interactive mode (default)
deepseek-code run "task" # Run a single task
deepseek-code init # Create DEEPSEEK.md template
deepseek-code history # Show conversation history
deepseek-code version # Show versiondeepseek-code run --help
Options:
-m, --model TEXT Model to use (deepseek-chat, deepseek-coder)
-t, --trust Trust mode: auto-approve safe operations
--yolo, --dangerously-skip-permissions
YOLO mode: skip ALL permission prompts
--max-turns INTEGER Maximum turns per task (default: 50)
-v, --verbose Verbose output
--no-context Don't load DEEPSEEK.md contextDeepSeek Code has a flexible permission system similar to Claude Code. Choose your level of automation:
deepseek-code- Read operations: Auto-approved (reading files, searching)
- Write operations: Asks for permission each time
- Shell commands: Asks for permission, blocks dangerous commands
- Option to "always" approve: Type
aoralwayswhen prompted to auto-approve similar operations
deepseek-code run --trust "your task"- Auto-approves safe write operations and shell commands
- Still asks for potentially dangerous operations
- Good for: Routine tasks in trusted projects
# Full flag (like Claude Code)
deepseek-code run --dangerously-skip-permissions "your task"
# Short alias
deepseek-code run --yolo "refactor entire codebase"- Skips ALL permission prompts - maximum speed
- Still blocks truly dangerous commands (
rm -rf /,sudo, fork bombs) - Shows warning banner when enabled:
⚠️ YOLO MODE ENABLED ⚠️ All permission prompts will be skipped! (Dangerous operations like rm -rf / are still blocked) - Good for: Experienced users, trusted environments, rapid iteration
| Mode | Flag | File Writes | Shell Commands | Dangerous Commands |
|---|---|---|---|---|
| Default | (none) | Ask | Ask | Blocked |
| Trust | --trust |
Auto | Auto (safe) | Blocked |
| YOLO | --yolo |
Auto | Auto | Blocked |
These are blocked in ALL modes for safety:
rm -rf /and similar destructive commandssudocommands- Fork bombs
- Direct disk writes (
dd if=/dev/zero of=/dev/sda) - Piping untrusted content to shell (
curl ... | bash)
DeepSeek Code comes with 6 built-in tools:
| Tool | Description | Permission |
|---|---|---|
read_file |
Read file contents with line numbers | Auto |
write_file |
Create or overwrite files | Ask |
edit_file |
Surgical string replacement (like sed) | Ask |
bash |
Run shell commands | Ask (dangerous blocked) |
glob |
Find files by pattern (**/*.py) |
Auto |
grep |
Search file contents with regex | Auto |
Create a DEEPSEEK.md file in your project root to give the AI project-specific knowledge:
deepseek-code init# DEEPSEEK.md
## Project Overview
FastAPI backend for user authentication.
## Key Commands
- `make test`: Run all tests
- `make lint`: Run linting
## Architecture
- `src/api/`: API routes
- `src/models/`: Database models
## Conventions
- Use type hints everywhere
- Write tests for new featuresdeepseek_code/
├── cli.py # Typer CLI entry point
├── agent.py # Core agent loop (the magic happens here)
├── llm.py # DeepSeek API client (OpenAI-compatible)
├── tools/ # Tool implementations
│ ├── base.py # Tool base class & registry
│ ├── file_tools.py # read, write, edit
│ ├── search_tools.py # glob, grep
│ └── bash_tool.py # Shell commands
├── permissions.py # Safety checks
├── context.py # DEEPSEEK.md loading
├── conversation.py # History management
└── ui.py # Rich terminal UI
The core is a simple agent loop:
while not done:
response = llm.chat(messages, tools)
if response.has_tool_calls:
for tool_call in response.tool_calls:
result = execute_tool(tool_call)
messages.append(result)
else:
done = True- Send messages to DeepSeek with available tools
- Execute tool calls if the AI requests them
- Feed results back to the AI
- Repeat until the AI responds without tool calls
This is the same pattern used by Claude Code, GPT-4 agents, and other AI coding assistants.
from deepseek_code.tools.base import Tool, ToolResult
class MyTool(Tool):
name = "my_tool"
description = "Does something useful"
permission_level = "ask" # or "auto" for safe operations
parameters = {
"input": {
"type": "string",
"description": "The input to process",
"required": True,
}
}
def execute(self, input: str) -> ToolResult:
# Your logic here
return ToolResult(success=True, output="Done!")Register it in tools/base.py:
def create_default_registry():
registry = ToolRegistry()
# ... existing tools ...
registry.register(MyTool())
return registry# Install dev dependencies
pip install -e ".[dev]"
# Run tests
pytest
# Format code
black deepseek_code/
ruff check deepseek_code/| Feature | DeepSeek Code | Claude Code | Cursor | GitHub Copilot |
|---|---|---|---|---|
| Open Source | Yes | No | No | No |
| Self-hostable | Yes | No | No | No |
| Cost | ~$0.14/M tokens | ~$15/M tokens | $20/mo | $10/mo |
| Customizable | Fully | No | Limited | No |
| Local LLM support | Planned | No | No | No |
- Streaming responses
- Sub-agent support for complex tasks
- Local LLM support (Ollama, llama.cpp)
- VS Code extension
- Multi-file editing
- Git integration tools
Contributions are welcome! Please feel free to submit a Pull Request.
MIT License - see LICENSE for details.
- Inspired by Claude Code architecture
- Powered by DeepSeek-V3
- Built with Typer and Rich
Star this repo if you find it useful!