Summary
Nick Saraev walks through a fully open-source method for using Claude Code—Anthropic’s terminal-based coding agent—without paying Anthropic directly, by intercepting the tool’s API calls and rerouting them through a local proxy server to cheaper model backends. The proxy, running on localhost:8082, can forward requests to NVIDIA NIM (free with an account), OpenRouter (models available from $0.14 per million input tokens), or a locally running Ollama instance that uses the user’s own GPU—all while preserving the exact same Claude Code CLI interface, commands, and UX.
The tutorial demonstrates the setup live by building a habit-tracking app called Habitual using DeepSeek V4 Flash as the backend, with Saraev claiming the project cost roughly 3 cents compared to an estimated $5–$10 using official Anthropic credits. He verifies the routing is working by showing the raw JSON request logs in OpenRouter, confirming DeepSeek V4 Flash is receiving the full Claude Code system prompt (~30,000 input tokens per request).
Saraev is upfront about the tradeoffs: DeepSeek V4 Flash produces between 23 and 59 tokens per second depending on the task, output quality lands around 80–90% of Claude Opus 4.7 for most coding work, and the architecture differences mean some tasks are faster and some slower than with Anthropic’s models. He frames the optimal strategy as using cheaper models for bulk refactoring while reserving frontier Opus-class models for the hardest reasoning tasks.
📺 Source: Nick Saraev · Published May 02, 2026
🏷️ Format: Tutorial Demo
