Local LLM + Claude Code: Cut AI Costs Without Cutting Quality
Split agent work by complexity: Claude Code for reasoning, a local Qwen model on a MacBook Air for mechanical tasks at zero token cost, with Ollama and oMLX benchmarked.
Henry Wang Cloud AI is expensive when you run it as an agent. A Claude Code session for a non-trivial feature can burn hundreds of thousands of tokens. Do that daily and the bill adds up.
My solution: split the work by complexity.
Claude Code handles the hard stuff — requirements, architecture, complex debugging, PR review. That’s where the frontier model earns its cost.
Local LLM via Pi handles the mechanical work — running tests and reporting failures, validating release checklists, triggering deployments. Repetitive, structured tasks that don’t need deep reasoning.
Token cost for those: zero.
I benchmarked this setup on a MacBook Air M5 32GB, running Qwen3.6 35B (4-bit) through two local runners — Ollama and oMLX.
For coding quality, I used Pi to autonomously complete two real tasks: write a Python data analysis script from scratch, and find and fix three bugs with unit tests.
Both runners scored 88–95 / 100. Correct output, clean code, tests passing.
oMLX was faster: 55 tok/s vs Ollama’s 44, and finished the bug-fixing task in 87 seconds vs 254. For a multi-step agent loop, that compounds quickly. Power draw: ~5W. The fanless MacBook Air never struggled.
The model uses ~25GB of 32GB RAM. Setup takes an afternoon.
Where I use Claude Code: feature planning, coding, PR review — anything that needs real reasoning.
Where I use Pi + local LLM: test runner agent, release checklist, deployment validation — anything mechanical and repeatable.
Local LLMs aren’t a replacement for Claude. They’re a way to stop paying Claude prices for tasks that don’t need Claude intelligence.
Full benchmark: mac-llm-bench report