中文

Quickstart

Deep Code is an open-source terminal AI coding assistant for the DeepSeek-V4 model, supporting deep thinking, reasoning effort control, and extend its capabilities with Skills and MCP.

Prerequisites

Before you start, make sure you have:

  • Node.js 22 or later
  • A DeepSeek API key

Install

Install Deep Code globally with npm:

npm install -g @vegamo/deepcode-cli

Check the installed version:

deepcode --version

Configure DeepSeek-V4

Deep Code recommends deepseek-v4-pro and also supports deepseek-v4-flash. Create ~/.deepcode/settings.json and add your DeepSeek model configuration:

{
  "env": {
    "MODEL": "deepseek-v4-pro",
    "BASE_URL": "https://api.deepseek.com",
    "API_KEY": "sk-..."
  },
  "thinkingEnabled": true,
  "reasoningEffort": "max"
}

Replace API_KEY with your DeepSeek API key.

Common fields:

FieldDescription
env.MODELDeepSeek model name, recommended deepseek-v4-pro
env.BASE_URLDeepSeek API endpoint, default https://api.deepseek.com
env.API_KEYDeepSeek API key
thinkingEnabledWhether to enable thinking mode
reasoningEffortReasoning effort, commonly "high" or "max"

You can also create .deepcode/settings.json inside a project to customize the model, permissions, or MCP settings for that project only.

For DeepSeek's official setup notes, see the Deep Code integration guide.

For all configuration options, see configuration_en.md.

Start

Open your project directory:

cd path/to/your/project
deepcode

Deep Code starts an interactive terminal UI in the current directory. Type a task and press Enter.

To start with an initial prompt:

deepcode -p "Summarize this project"

Try These First

Start with a read-only task:

Summarize this repository and explain how to run it.
Find the main entry points and explain the request flow.

Then try a coding task:

Add a unit test for the login validation logic.
Run the test suite and fix the failing tests.

You can also ask for a plan first:

Before editing files, propose a plan for adding pagination to the user list.

Basic Controls

ActionKey
Send messageEnter
Insert a newlineShift+Enter or Ctrl+J
Interrupt the current responseEsc
Paste an imageCtrl+V
QuitPress Ctrl+D twice, or use /exit

Slash Commands

Type / in the input box to open the command menu.

CommandAction
/newStart a new conversation
/resumeChoose a previous conversation to continue
/continueContinue the current conversation or resume the latest one
/modelSwitch model, thinking mode, and reasoning effort
/initCreate an AGENTS.md instruction file for the current project
/skillsShow available Agent Skills
/mcpShow MCP server status and available tools
/undoRestore code and/or conversation to an earlier point
/rawChange the display mode
/exitQuit Deep Code

Add Project Instructions

Run this inside a project:

/init

Deep Code helps create AGENTS.md. Use it to record project conventions, such as:

  • How to install dependencies and run tests
  • Code style and contribution expectations
  • Important directory notes
  • Checks to run before or after editing code

Deep Code automatically uses these instructions when working in the project.

Use Skills

Agent Skills are reusable workflows, such as code review, release checks, documentation generation, or framework-specific development steps.

List available skills:

/skills

You can also type / and choose a skill from the menu.

For more details, see agent-skills_en.md.

Connect External Tools

Use MCP to connect Deep Code to GitHub, browsers, databases, or other services.

After configuring MCP, run:

/mcp

This shows connected MCP servers and available tools.

For setup instructions, see mcp_en.md.

Permissions and Safety

Deep Code may read files, edit code, or run commands. You can configure which actions are allowed automatically, which require confirmation, and which are denied.

Deep Code supports YOLO mode by default, so it can smoothly read and write files, run commands, and continue common coding tasks. If you prefer a more cautious setup, use strict permissions so Deep Code asks before higher-risk actions.

For details, see permission_en.md.

Task Completion Notifications

Deep Code can run a notification script when a task finishes, such as sending a Slack message, Feishu message, system notification, or terminal alert.

For examples, see notify_en.md.

Next Steps