DeepSeek GitHub Copilot Setup: Chat Extension and CLI

Introduction

Search DeepSeek GitHub Copilot setup, and many people blur two paths: install an extension into Copilot Chat, or feed environment variables to Copilot CLI. Both can run DeepSeek — the config is not the same.

Short pick: stay inside the Copilot Chat UI with Agent / tools / MCP → install “DeepSeek V4 for Copilot Chat”; run copilot in a terminal → BYOK, and you must set the provider to anthropic. This site is an independent community; steps follow official Integrate with GitHub Copilot and Integrate with Copilot CLI. Models and rates: pricing.

I. Pick a Path: Copilot Chat Extension vs Copilot CLI

You want	Use	Key point
Stay in Copilot Chat UI; keep Agent / Skills / MCP	VS Code extension DeepSeek V4 for Copilot Chat	Set API Key via Command Palette; pick Pro / Flash in the model picker
Run copilot in a terminal with your own key	Copilot CLI BYOK	COPILOT_PROVIDER_TYPE=anthropic + DeepSeek Anthropic base_url
OpenAI Codex client ecosystem	Not this article	See DeepSeek Codex setup
Anthropic Claude Code	Not this article	See Claude Code DeepSeek

Both Copilot paths need a key from the DeepSeek Platform (usually sk- prefix) and bill against your DeepSeek API balance — free web chat is unrelated. Overview: DeepSeek API.

II. Path A: DeepSeek V4 for Copilot Chat (VS Code)

Prerequisites: VS Code 1.116+; a GitHub Copilot subscription (Free / Pro / Enterprise — free tier works).

1. Install DeepSeek V4 for Copilot Chat (or from the GitHub repo). When official docs say “install from the Github repo,” use the current Marketplace / repo entry.
2. Run DeepSeek: Set API Key from the Command Palette and paste the key. It is stored in the OS keychain, not on disk.
3. Open Copilot Chat (Cmd+Shift+I / Ctrl+Shift+I), open the model picker, choose DeepSeek V4 Pro or DeepSeek V4 Flash.
4. Chat away — Agent mode, tool calling, Skills, and MCP stay on Copilot’s surface; inference is DeepSeek.

Thinking effort: the gear next to the model offers None (fastest, no reasoning) / High (default) / Max (hard tasks). Thinking tokens bill as output — start at High, raise to Max when needed. Rules: DeepSeek thinking mode.

Vision: official notes treat DeepSeek V4 as text-side here; drop a screenshot and the extension proxies through another installed Copilot model (Claude, GPT-4o, …) for a description, then sends that to DeepSeek. Use DeepSeek: Set Vision Proxy Model to pick the describer. For true image blocks on the API, see How to send images with the DeepSeek API (deepseek-flash).

III. Path B: GitHub Copilot CLI (anthropic required)

Prerequisites: Node.js 22+; install globally:

npm install -g @github/copilot

Then set env vars — Linux / macOS:

export COPILOT_PROVIDER_TYPE=anthropic
export COPILOT_PROVIDER_BASE_URL=https://api.deepseek.com/anthropic
export COPILOT_PROVIDER_API_KEY=sk-your-deepseek-api-key
export COPILOT_MODEL=deepseek-v4-pro

Windows (PowerShell):

$env:COPILOT_PROVIDER_TYPE="anthropic"
$env:COPILOT_PROVIDER_BASE_URL="https://api.deepseek.com/anthropic"
$env:COPILOT_PROVIDER_API_KEY="sk-your-deepseek-api-key"
$env:COPILOT_MODEL="deepseek-v4-pro"

Switch COPILOT_MODEL to deepseek-v4-flash when you want Flash. Run copilot. Full Agent mode, tool calling, and MCP are supported.

Hard rule: do not set COPILOT_PROVIDER_TYPE to openai. Official docs: the openai type triggers 400 — The reasoning_content in the thinking mode must be passed back to the API. DeepSeek requires echoing thinking content; Copilot CLI’s OpenAI integration cannot. Use the Anthropic Messages endpoint instead.

Optional token caps (recommended when Pro is not in the CLI’s built-in catalog):

export COPILOT_PROVIDER_MAX_PROMPT_TOKENS=840000
export COPILOT_PROVIDER_MAX_OUTPUT_TOKENS=128000

More variables: copilot help providers. COPILOT_OFFLINE=true only blocks GitHub-side calls; prompts still go to api.deepseek.com — not “fully offline.”

IV. Flash vs Pro and Peak Pricing

Scene	Tip
Daily edits, explain code, high-frequency small tasks	Flash + thinking High (or None/High in Chat)
Hard refactors, long agent chains	Pro; Max when needed
Screenshots in Chat	Chat extension + Vision Proxy; or API-side deepseek-flash for real vision

Peak windows: DeepSeek API peak hours. 429s: DeepSeek API 429. Copilot sessions are tool-heavy and long-output — you pay DeepSeek API, not “Copilot subscription covers DeepSeek.”

FAQ

How do I choose vs Codex, Claude Code, or Harness?
Stay on GitHub Copilot UI / CLI → this article. OpenAI Codex clients → Codex setup. Claude Code → Claude Code wiring. DeepSeek’s own agent runtime → Harness. They can coexist; pick by client.

Can Chat and CLI share one key?
Yes. The key belongs to your DeepSeek account — watch concurrency and balance when several heavy agent sessions run at once.

Do I still need LiteLLM or a local proxy?
On the official Chat extension and CLI Anthropic BYOK paths, you usually do not need a separate protocol-translation layer. If an old guide only mentions proxies, verify these two official paths first.

Summary

Shortest path for DeepSeek GitHub Copilot setup: Chat → install the extension → Set API Key → pick Pro/Flash; CLI → anthropic provider + https://api.deepseek.com/anthropic → set model and run. Open both official guides before changing env; never use the openai provider type on CLI.

Further reading

• DeepSeek API
• DeepSeek Codex setup
• Claude Code DeepSeek
• What is DeepSeek Harness
• DeepSeek thinking mode
• DeepSeek API 429