DeepSeek V4 smuggled into Copilot Chat's model picker

What it does

This extension registers DeepSeek V4 Pro and Flash as options inside GitHub Copilot Chat’s native model picker. You chat with DeepSeek through Copilot’s familiar sidebar, keeping agent mode, tool calling, custom instructions, and the rest of the Copilot stack intact. You bring your own DeepSeek API key, which the extension stores in the OS keychain rather than on disk.

The interesting bit

DeepSeek V4 is text-only, so the extension quietly intercepts any image you drop into chat, routes it through another installed Copilot model for a description, and feeds that text back to DeepSeek. It is essentially a vision compatibility bridge that lets a text-only model participate in multimodal conversations without a separate UI or workflow.

Key highlights

• Plugs into Copilot’s native provider API; no new chat window or sidebar to learn
• Supports DeepSeek V4’s thinking mode with native reasoning effort controls (None / High / Max)
• Proxies vision requests through other Copilot models so image chat works out of the box
• API key lives in VS Code SecretStorage (OS keychain), never in settings.json
• Zero runtime dependencies; pure VS Code API with no local proxy server to manage

Caveats

• Relies on non-public Copilot Chat APIs that may break with future VS Code updates
• Vision is proxied through another model rather than handled natively by DeepSeek
• The experimental stabilizeToolList setting can inflate input tokens if enabled with too many tools

Verdict

Worth installing if you like Copilot’s agent mode and polished UI but want DeepSeek’s pricing and 1M-token context window. Skip it if you are allergic to extensions that hook into undocumented APIs or need native multimodal reasoning without workarounds.