While waiting for local models to get better, you might be like me: using a low tier of a subscription-based LLM service. In my case, that is Claude Pro. A hallmark of those services is that there are some form of limits for how much work you can do within the limits of the subscription level, before you get hit with some sort of overage fees. In the case of Claude Pro and Max, you have 5-hour and 7-day "sessions", with certain levels of usage within those sessions. Go past that amount of usage, and you start "paying full freight" for your LLM invocations — that is annoying at best and expensive at worst.

So, it behooves you to have some idea of how well you are doing with respect to those session limits. Not only does your usage vary, but so do the limits, ranging from peak-versus-off-peak distinctions to LLM providers changing the limits based on changing requirements of their businesses.

If you are doing interactive LLM work, such as agentic coding, ideally your agent harness keeps you informed of your usage and limits. While you could just keep hammering the /usage command (or its equivalent), it would be nicer if your harness displayed this information automatically, like a car dashboard displays speed and battery/fuel levels.

Alas, Claude Code does not offer this by default.

What Claude Code does offer is the ability to customize the status lines at the bottom of the terminal output. Those can be controlled by executables that receive some JSON and output what to display. Lots of status line implementations are available, and some will show your usage.

I have settled on claude-pace. It is a single ~300-line Bash script, relying on jq to handle JSON parsing. While there is a Claude plugin you can use, I elected to install the Bash script manually and update it from time to time. What I like about it, besides its small size and semi-readable code, is that it shows how quickly you are exhausting the session. Knowing that I have used 62% of a 5-hour session limit is one thing; knowing that at my current pace I am likely to run out before the end of the session is another.

This status line is only really meaningful for Claude Pro and (presumably) Max subscriptions. I installed it on a machine that I am using for a client with their Enterprise subscription, and while the status line works without crashing or something, Enterprise limits seem sufficiently different that claude-pace is not useful.

Monitoring usage is important, whether your objective is to minimize your frontier LLM usage or just to avoid paying extra. claude-pace is a nice option for Claude Code, but no matter what your agent harness is, you might want a way to track your usage, if not your pace.

koverGate and detektifier, my Gradle plugins for helping agents use Kover and Detekt, are each up to version 0.1.3. The only significant changes were dependency updates.