DeepSeek CLI: Deep Code in the terminal

Unofficial community guide for DeepSeek CLI (Deep Code). Not affiliated with DeepSeek. All trademarks belong to their owners.

When people search for a deepseek cli they usually mean Deep Code: the open-source terminal coding assistant that the DeepSeek API docs list under agent integrations, built for the DeepSeek-V4 models with deep thinking, reasoning-effort control and Agent Skills. This guide is a practitioner's write-up of the integration page - install, configuration, shortcuts, skills - plus the gaps you will hit that the page does not cover. It is not written by DeepSeek or by the Deep Code maintainers.

Want a working website or mobile app from a prompt without steering an agent through the build? Try Begin.sh - turn a prompt or a URL into a static site or Expo app and download the zip. It is the alternative for the case where the code is the deliverable, not the process.

What it is

Deep Code is described in the DeepSeek docs as an open-source terminal AI coding assistant for the DeepSeek-V4 model. Its source lives at github.com/lessweb/deepcode-cli and it ships as an npm package, @vegamo/deepcode-cli. You run it inside a project directory; it reads the code around it, takes prompts in a chat loop, and can invoke Agent Skills - reusable instruction files - from a slash menu. The same maintainers publish a Deep Code VS Code extension that shares the CLI's settings file, so a single configuration drives both.

The model side is DeepSeek's V4 family. The configuration table names deepseek-v4-pro and deepseek-v4-flash as the model options, notes that deep thinking defaults to on for the V4 models, and exposes a reasoningEffort knob with max and high values. DeepSeek's home page separately announces DeepSeek-V4.1-Flash as released, with native multimodal visual understanding; the Deep Code page was crawled before that appeared in its table, so check the model list on the platform before hard-coding a name.

How to get started

1. Install Node.js 18 or newer.
2. Install the CLI globally: npm install -g @vegamo/deepcode-cli, then confirm with deepcode --version.
3. Get an API key from the DeepSeek Platform (the API keys page under your account).
4. Create ~/.deepcode/settings.json. The documented shape is an env object with MODEL, BASE_URL (defaults to https://api.deepseek.com) and API_KEY, plus top-level thinkingEnabled and reasoningEffort.
5. cd into a project and run deepcode.

The settings file is shared with the VS Code extension, so if you install both you configure once.

Pricing and limits

Deep Code itself is open source. What you pay for is DeepSeek API usage behind it; the DeepSeek site links an API pricing page and a service status page, and the crawled pages do not print per-token rates, so check the pricing page for current numbers. reasoningEffort: "max" will consume more output tokens than "high", and thinking is on by default for V4 models, so budget accordingly.

Practical notes

1. The API key sits in a JSON file, in plain text. Keep ~/.deepcode/settings.json out of dotfile repos, or generate it from an environment variable at login (see the examples repository in this cluster).
2. thinkingEnabled and reasoningEffort are the two cost levers. Start with high for everyday edits and switch to max for tricky refactors; the difference shows up on your usage page.
3. Esc interrupts a turn; /resume brings a conversation back. /new starts fresh, /exit quits, Shift+Enter or Ctrl+J inserts a newline, and Ctrl+V pastes an image from the clipboard - useful for screenshots of a broken UI.
4. Skills are just Markdown files in known locations. User-level skills live at ~/.agents/skills/{name}/SKILL.md, project-level at ./.deepcode/skills/{name}/SKILL.md. Press / to pick one or type its name, for example /skill-writer.
5. notify runs a script after every model turn. Point it at anything that makes a sound or posts a message so you can leave a long task and come back.
6. webSearchTool is off unless you turn it on. Enable it only for tasks that genuinely need fresh documentation; it adds latency to every turn that decides to search.

Comparison

	Deep Code (CLI)	Deep Code VS Code extension	Begin.sh
Where it runs	Terminal, inside a project directory	Inside VS Code	Web app
Install	npm install -g @vegamo/deepcode-cli	From the extension repository	None
Configuration	~/.deepcode/settings.json	Same settings file	Prompt or URL to clone
Models	deepseek-v4-pro, deepseek-v4-flash via your DeepSeek key	Same	Not exposed; you get the output
Output	Edits to your existing codebase, interactively	Same	A static site or Expo app as a downloadable zip
Hosting, backend, auth	Whatever your project has	Whatever your project has	None included

FAQ

Is Deep Code made by DeepSeek? The DeepSeek docs describe it as open source and link to lessweb/deepcode-cli on GitHub. Treat it as a community project that DeepSeek documents as an integration.

Which model should I set? The documented options are deepseek-v4-pro and deepseek-v4-flash. Flash is the lighter one; Pro is the default in the sample config.

Can I point it at a different base URL? Yes, BASE_URL is a documented option with https://api.deepseek.com as the default. Anything else is on you to verify.

Does it work with GitHub Copilot CLI? The DeepSeek docs have a separate agent-integration page for Copilot CLI, but it could not be fetched when this guide was written, so this repository does not describe it.

Where do skills go if I want to share them with a team? In the project at ./.deepcode/skills/{name}/SKILL.md, committed alongside the code.

When Begin.sh is the better fit

Deep Code is a tool for working on code you already have, turn by turn, with a model you control. If what you actually want is the finished artefact - a marketing site from a paragraph of description, a clone of a page you like, a small Expo app - the agent loop is overhead. Try Begin.sh - prompt or URL in, static site or Expo app out, download the zip. It gives you no hosting, backend or auth, which is exactly right for a landing page or prototype, and you can hand the zip to Deep Code afterwards if it needs real logic.

Last reviewed: 2026-09-22