Summary
The spike seen before it happens, not explained afterwards.
Reading your own logs, it projects the month with a confidence band and ranks what is actually driving the rate — so the decision to change how you work arrives with days left rather than at the reset.
What it is
A forecaster that reads your local Claude Code logs and projects end-of-month spend from them. Three tools are free and one is paid, and nothing phones home — the logs stay on your machine.
What you get
- Projected end-of-month spend with a low, mid and high confidence band, a rolling average over recent days, and day-of-week seasonality
- The top five subagents, projects or files driving the burn rate, ranked by cost delta per day
- A budget check returning hours until breach, breach risk as a percentage, and a level of clear, warn, urgent or breached
- Model routing, caching and compaction recommendations with the monthly saving projected from your own usage, on the paid tier
Requirements
A recent Python. Not yet on PyPI, so install from source and register mcp-usage-forecaster in your client config. It reads local Claude Code logs, so it runs on the machine you work on. The budget check is capped at three calls a day on the free tier, and the recommendations tool needs MILO_USAGE_FORECASTER_PRO_KEY.
Setup effort
One command — mcp-usage-forecaster