Everything a directory asks for.

Notices when you're cooked before you do.

A load-aware companion for people who live in Claude.

Tempo watches how long you've actually been in Claude and how the session is going, then offers one 40-second reset at a pause. Fully local, never interrupts.

Tempo is a load-aware companion for people who work long hours with Claude — chat, Cowork, Claude in Chrome or Claude Code. It reads the shape of a session and never its words: how many hours you've actually been in Claude today, how long since a real break, how many codebases you're holding at once, plus friction signals like retry spirals and error rates above your own baseline. When the pattern says your nervous system has left the building, it offers exactly one thing — a 40-second reset, at a pause, citing the signal that earned it. At most twice a day, and a single no silences it. Everything stays on your machine: no account, no telemetry, no network calls anywhere in the codebase.

Most tooling for people who work with AI assistants optimises throughput. Tempo measures cost. It installs into Claude Code as a set of hooks and a statusline, then scores your session load from derived signals only — never prompt text, file contents, paths or command strings. Two families feed the score. Friction signals spike when something is going wrong: clusters of retries on the same target, error rates above your own two-week baseline, prompts collapsing from sentences into fragments. Dose signals accumulate whether or not anything is wrong: active minutes in Claude today with idle gaps excluded, minutes since your last real break, prompt velocity against your own baseline rate, how many separate codebases are in play, and consecutive days without a lighter one. Enough dose alone earns an offer, because an eight-hour day doesn't have to go badly to have cost you something. The statusline glyph warms from green to amber to red as load rises, and shows your running total for the day. At the top of the scale, and only in a seam — after a turn completes, when nothing is running — Tempo offers one 40-second nervous-system practice from The 40-Second Reset, citing the signal that earned it. Maximum two offers a day, ninety minutes apart, and a decline silences it. Nothing is gamified: no streaks, no praise, no score to chase. A weekly summary built from your own logs shows your best hour, your worst hour, and whether the resets measurably helped — including when they didn't. There are no network calls anywhere in the codebase, no account and no server, and rm -rf ~/.tempo is a complete erasure. For Claude chat, Cowork and Claude in Chrome, the same rules ship as a single skill file with no requirements at all.

Agentic coding removed the natural pauses from the work. Waiting on a compile used to be a break; waiting on an agent isn't, because you're reading, steering and deciding the whole time. So the hours stack up with nothing in them that stops you. Tempo is instrumentation for that: it measures the sessions you're actually having and speaks once, quietly, when the pattern has earned it.