On this page

Cloud metrics

For Claude Code, Codex, Cursor, and similar services, track percent used, reset time, credits, and pay-as-you-go spend. These numbers answer whether the account can continue and what the next request may cost.

Also track the model and task type. A cheap, fast model and a premium reasoning model can consume the same plan at very different rates.

Local metrics

For Ollama and LM Studio, track RAM or VRAM, loaded model size, context used, tokens per second, queue state, and model load time. There may be no per-request invoice, but slow or overloaded hardware is still a real operating cost.

Local usage is also affected by electricity and the opportunity cost of tying up the machine. Those factors matter most for long-running or parallel workloads.

Keep your usage visible while you work. Get Super Notchy for $27 ↗

A practical routing rule

Use local models for private, repetitive, or lightweight work when the quality is sufficient. Use cloud models where stronger reasoning, larger hosted capacity, or integrated tools earn the additional usage.

A combined dashboard makes the tradeoff visible. Super Notchy places cloud allowances and local runtime signals in the same notch, while keeping the units honest instead of pretending they are one universal percentage.

Sources and further reading