DeepSeek V4 + Claude Code: hybrid coding setup

This tutorial pairs DeepSeek V4 with Claude Code through a local Anthropic-compatible proxy. The demonstrated setup sends requests to DeepSeek's cloud API; running the proxy on your computer does not make model inference local. The speaker calls the model V4, but the saved prompt uses deepseek-chat or equivalent, leaving its exact runtime version unverified. The speaker describes local deployment as hardware intensive and cautions against sending sensitive data through the cloud API.

The setup requires Claude Code and a DeepSeek API key. Antigravity configures the free-claude-code repository and creates launch scripts for the proxy and two separate coding sessions. The proxy must start before the routed session. One session uses DeepSeek for scripts, unit tests and scaffolding, while the native Claude session uses Opus 4.7 for more complex work.

The coding assistant demo builds an AI agent dashboard in a shared project folder. DeepSeek handles the initial structure and mock data; Opus works on layout, interactions and component quality. The result uses mock data without a connected API, so it does not establish that the dashboard is ready for production.

The speaker recommends reserving stronger models for serious web development, code reviews and security audits. He reports about 15 cents in spending while also using a paid Claude plan; the accounting scope is unspecified. No controlled cost comparison establishes what other projects would cost.