Deepseek Family for Github Copilot Chat
Bring DeepSeek V4.1 Flash and DeepSeek V4-Pro into GitHub Copilot Chat — with visible chain-of-thought, tool calling, native image input on DeepSeek Flash, and a live account-balance status bar.
Install the Extension from VSCode Marketplace or search DeepSeek Family for GitHub Copilot Chat in VSCode Extension Store.
What's new in 1.0.3 — switched the built-in Flash entry to the new official deepseek-flash model, added native V4.1 Flash image-input support, restored the official None / Low / High / Max thinking-effort options, and kept retired Flash ids working through compatibility routing. See the changelog.
This extension is modeled after glm-for-copilot and adapted for the DeepSeek OpenAI-compatible API. Huge shout out to @KiwiGaze for his foundamental work and generous licensing.
Highlights
- 🧠 Selectable thinking effort — every model exposes a None / Low / High / Max selector in the Copilot Chat model picker. DeepSeek's
reasoning_content is forwarded to VS Code's thinking UI.
- 🚀 Current official model lineup — the picker exposes
deepseek-flash as DeepSeek V4.1 Flash and keeps deepseek-v4-pro available as the current Pro endpoint.
- 🖼️ Native image input on DeepSeek Flash — image attachments are serialized into DeepSeek's official Chat Completions content-block format and sent to
deepseek-flash.
- 🔁 Compatibility for retired Flash ids — if you pinned
deepseek-v4-flash or deepseek-v4-flash-vision-exp, the extension maps them to deepseek-flash so old selections keep working.
- 💹 Peak/off-peak pricing in the usage panel — the usage detail view now lists per-1M-token input/output prices for each model. Chinese UI shows the official CNY table from the Chinese pricing page; English UI shows the official USD table from the English pricing page.
- 🛠️ Full Copilot Chat integration — tool calling, multi-turn reasoning preservation (passes
reasoning_content back across tool-call turns to avoid the 400), and proper tool_choice handling.
- 🔐 Bring-your-own-key — your key lives in VS Code's SecretStorage (the OS keychain). Never written to
settings.json or Git.
- 💰 Live account balance — a status-bar item polls
GET /user/balance and surfaces your remaining credit; turns red when is_available: false.
- 🌐 Single endpoint — DeepSeek exposes one global base URL (
https://api.deepseek.com); no region/apiMode gymnastics needed.
Quick start
- Install the extension.
- Create a key at https://platform.deepseek.com/api_keys.
- Run DeepSeek: Set API Key from the Command Palette and paste the key.
- Open Copilot Chat, click the model picker, and pick DeepSeek V4.1 Flash or DeepSeek V4-Pro.
- If you want to attach screenshots or other images, use DeepSeek V4.1 Flash.
- (Optional) Use the per-model Thinking Effort selector to choose
None, Low, High, or Max.
Settings
| Setting |
Default |
Description |
deepseek-copilot.baseUrl |
"" |
Override the DeepSeek API base URL (empty = official endpoint). |
deepseek-copilot.maxTokens |
0 |
Max output tokens per request (0 = API default). |
deepseek-copilot.maxRetries |
3 |
Auto retries on HTTP 429 / 5xx (exponential backoff + Retry-After). |
deepseek-copilot.modelIdOverrides |
{} |
Remap picker model ids to API model ids. |
deepseek-copilot.customModels |
[] |
Add your own model entries to the picker. |
deepseek-copilot.debugLogging |
false |
Verbose logs to the DeepSeek output channel. |
deepseek-copilot.showUsageStatusBar |
true |
Show the balance status bar. |
deepseek-copilot.usageRefreshIntervalMinutes |
5 |
Refresh interval for the balance status bar (1–1440). |
Commands
| Command |
Description |
DeepSeek: Set API Key |
Save your DeepSeek key to SecretStorage. |
DeepSeek: Get API Key |
Open platform.deepseek.com/api_keys. |
DeepSeek: Clear API Key |
Remove the stored key. |
DeepSeek: Open Settings |
Jump to the extension's settings. |
DeepSeek: Show Logs |
Reveal the DeepSeek output channel. |
DeepSeek: Refresh Usage |
Force-refresh the balance status bar. |
DeepSeek: Show Usage Details |
Open the balance detail webview. |
How thinking effort works
| Picker value |
Wire body |
| None |
thinking: { type: "disabled" }, reasoning_effort: "none" |
| Low |
thinking: { type: "enabled" }, reasoning_effort: "low" |
| High |
thinking: { type: "enabled" }, reasoning_effort: "high" |
| Max |
thinking: { type: "enabled" }, reasoning_effort: "max" |
DeepSeek's default is High. Picking Max is recommended for hard coding and agentic tasks
(as Claude Code and OpenCode do automatically). Pick None to disable chain-of-thought entirely.
According to the current DeepSeek docs, legacy minimal normalizes to Low, and medium / xhigh
both normalize to High.
Vision model notes
- DeepSeek V4.1 Flash is the current official image-capable model.
- Images are sent through
POST /chat/completions using DeepSeek's official content block array with inline image_url data URLs.
- DeepSeek only accepts images in
user messages, not system, assistant, or tool messages.
- Supported image formats are JPEG, PNG, GIF, and WebP.
- DeepSeek's retired ids
deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily routed by the backend to deepseek-flash.
- According to the latest pricing page,
deepseek-v4-pro does not expose image understanding as a first-class capability; DeepSeek plans to route it to V4.1 Flash after 2026-09-14 12:00 Beijing time until V4.1 Pro is released.
Pricing in the extension
- Open DeepSeek: Show Usage Details to see the current balance plus official per-model pricing.
- The panel lists off-peak and peak unit prices for input (cache hit), input (cache miss), and output, all per 1M tokens.
- In Chinese UI, prices are shown in CNY using the official Chinese pricing page.
- In English UI, prices are shown in USD using the official English pricing page.
- DeepSeek updates pricing frequently, so the panel includes a source note and routing note for retired ids.
See docs/deepseek-api.md for the full API reference.
Development
pnpm install
pnpm run compile # build to out/
pnpm run watch # rebuild on save
pnpm run typecheck # tsc --noEmit
pnpm run lint # eslint
pnpm test # vitest run
pnpm run package # produce a .vsix in dist/
Press F5 in VS Code to launch an Extension Development Host with the extension loaded.
License
MIT
