Context, quota, cache and cost in the status line. Read from stdin, computed in Go, rendered before you notice. No network calls, no dependencies.
[Opus 4.6] | commander@ai-scream.ai | main* | Out:1K | $185.4 | 91h12m ████░░░░░░ 41% ( 82K/200K) | █████████░ 94% (4h37m/5h) | ███████░░░ 79% (2d21h/7d) Δ+3.4K/-1.3K | Cache:99%(W:0K/R:82K) | Wait:6% | Cost:$0.03/m | Insert | v2.1.39 Bash(5) Edit(1) Read(3) | ▶explore
Every segment is one metric with its own color thresholds. Turn any of them off.
A ten-character bar with absolute tokens, not just a percentage. Formats as K or M, so a 1M window reads the same way.
The 5-hour and 7-day limits with time to reset, taken from the rate_limits field Claude Code already sends.
How much of the last call hit the prompt cache, and what share of the session was spent waiting on the API.
Dollars per minute next to the session total. Fifteen thresholds decide when each color changes.
When context crosses the danger threshold, Howl collapses to two dense lines: tokens left, time left, and the numbers that still matter. Use the controls above the HUD to switch presets and modes, or click a preset here.
Howl ships as a Claude Code plugin from its own marketplace. Setup downloads the binary for your OS and writes the status line setting for you.
/plugin marketplace add ai-screams/howl /plugin install howl@ai-screams-howl /howl:setup
Howl is free and MIT licensed. If it earns a place in your terminal, sponsoring on GitHub helps keep it maintained.