LLM Cost Tracer (Claude Skill)

Adds per-call tracing to every LLM request (model, tokens, cache, latency, USD, errors) in a JSONL file, with a spend and provider-health report.

Drop-in wrappers for Python and Node that record the provider's own usage numbers, including cache reads and writes, plus failures. A report shows spend by feature and model, the most expensive calls, and per-provider health. Unknown models are recorded with usd: null rather than a guessed price. Where it comes from: The first provider-health run on my site showed most Anthropic calls failing on credit balance. Visitors never noticed, because the fallback answered.

Knowing what each AI feature costs and noticing a dead key that a fallback is quietly covering for.

# Claude Code (personal, all projects)
unzip llm-cost-tracer.zip -d ~/.claude/skills/

# Claude Code (one project)
unzip llm-cost-tracer.zip -d .claude/skills/

# Claude.ai / Claude desktop
Settings → Capabilities → Skills → Upload skill → choose the zip