Local models don't send you an invoice, so it's tempting to stop measuring anything. Here's why I log, manage, and visualise token usage anyway — and the ties back to the context-window bug that bit me a few posts ago.
I pointed my offline coding agent — Ollama, Cline, and Qwen2.5-Coder:14b — at a real backlog from Phonics Journey. Nine Dependabot PRs and a pile of words falling back to a sparkle emoji.
I wanted a coding agent that doesn't phone home. Here's how I got Qwen2.5-Coder:14b running entirely on a 16GB MacBook Air M4, using Ollama and Jan, wired into VS Code as an actual agent.