lessweb/deepcode-cli

Deep Code CLI – a terminal AI coding assistant

What it is – A Node‑based command‑line tool that lets you chat with the DeepSeek‑v4 family of large language models (LLM) directly from any project folder. It is tuned for DeepSeek, supports the model’s thinking mode and context‑caching, and can be extended with user‑written agent skills.

Installation

npm install -g @vegamo/deepcode-cli   # global install

After installation, run deepcode inside a project directory to start an interactive session.

Configuration Create ~/.deepcode/settings.json (shared with the VS Code extension). A minimal example:

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

The docs (docs/configuration.md) describe multi‑level overrides, environment‑variable support, and additional flags such as filesApiEnabled, webSearchTool, notify, and permission policies.

Core features

Feature	What it does
Agent Skills	Load custom command modules from ./.deepcode/skills/, ./.agents/skills/, ~/.deepcode/skills/ or ~/.agents/skills/. Skills can add new commands, tools, or data sources.
DeepSeek‑specific optimisations	Uses DeepSeek KV‑cache to reduce token cost, enables the official thinking mode and lets you control reasoning effort (max, medium, …).
Slash commands & hotkeys	/new, /resume, /fork, /model, /raw, /skills, /mcp, /undo, etc., plus Enter/Shift‑Enter, Ctrl‑V (paste image), Esc (interrupt), Ctrl‑D (exit).
Multimodal support	When a vision‑capable model (e.g., deepseek-v4-flash-vision-exp) is selected, you can paste or load images; non‑vision models fall back to an UnderstandImage tool.
MCP (Model Context Protocol)	Connect to external services (GitHub, browsers, databases) via configurable MCP servers; view status with /mcp.
Permission system	Fine‑grained control over file, shell, and network actions (allow, ask, deny) via the permissions field.
Notifications	After a task finishes, a user‑provided script (e.g., Slack webhook) can be run automatically (notify field).
Cross‑model compatibility	Besides DeepSeek, any OpenAI‑compatible endpoint can be used (e.g., Volces Coding Plan).

Typical workflow

1. deepcode → start a REPL.
2. Type a prompt or use a slash command (/new, /model deepseek-v4-flash …).
3. The assistant replies, optionally showing a plan or raw output.
4. Use /undo to revert code changes, /skills to list custom extensions, or /mcp to invoke external tools.
5. When finished, the configured notification script runs.

Extending the tool

• Write a skill as a small Node module placed in one of the skill directories; it will be discovered automatically.
• Add custom web‑search or file‑upload scripts via webSearchTool or filesApiEnabled.
• Configure additional MCP servers in settings.json to let the assistant browse the web, query databases, etc.

Related projects

• Deep Code VS Code extension – shares the same settings file and provides the same AI features inside the editor.
• deepcode-qrcode‑benchmark – a benchmark showing Deep Code + DeepSeek outperforming Claude Code on a Python QR‑code task.

Development & contribution

git clone https://github.com/lessweb/deepcode-cli.git
cd deepcode-cli
npm install
npm test          # run tests
npm run build      # lint, type‑check, compile
npm link           # link globally for local testing
npm run build:vscode   # build the VSCode extension

Before submitting a PR, run npm run check (type‑check + lint) and format the code with npm run format.

License – MIT.

---

All information above is taken directly from the repository’s README; no additional features have been inferred.