Summary
Nate Herk walks through two practical methods for running Claude Code without accumulating Anthropic API costs: swapping in locally hosted open-weight models via Ollama, and routing requests through Open Router’s free model tier. The video opens with a clean conceptual frame — Claude Code is the orchestration harness (the ‘car’), and the language model is the engine, which can be replaced independently.
To justify the substitution, Herk references SWE-bench verified leaderboard scores showing that several current open-weight models now outperform Claude Sonnet 3.7, including Qwen variants available through Ollama. The Ollama integration demo is honest about the tradeoffs: local models don’t surface tool call steps during inference the way cloud Claude does, response times can stretch to four minutes or more for project-level context scans, and context window limits in Ollama sometimes silently truncate large codebases without warning. The Open Router path offers faster inference and broader model selection while still operating within free tier limits for most developers.
The tutorial covers the full configuration path for both methods — setting a custom base URL, handling authentication quirks, and understanding where open-weight models tend to break Claude Code’s expected JSON tool-calling protocol. Developers who want to experiment with Claude Code’s agentic scaffolding without the per-token cost will find both approaches documented with enough specificity to replicate.
📺 Source: Nate Herk | AI Automation · Published April 04, 2026
🏷️ Format: Tutorial Demo
