What it does
Claude Code is Anthropic's AI agent that works on your computer. With a Claude subscription you do not pay per job: you have two limits, one for each 5 hours and one for the week, and when one runs out you wait. Claude Code does not tell you what a job will use before it starts, so you find out only afterwards: a big refactor or a long research can hit the 5-hour limit halfway through, or use up most of your week.
Quotient is a free plugin that gives you the price first.
- You see the price first. Above a size you set, Claude stops and shows what the job would cost, how long it would take and how much of your limits it would use. You choose.
- Big jobs fit. A job too big for today can run in pieces (installments), at night if you like, each with its own spending cap.
- It learns. Quotient measures what each job really cost, and the next quote learns from the difference.
It is an estimate, not a promise. On real jobs the typical miss is still around 40%, and sometimes much more: the real numbers are below, including the bad ones.
In pictures
Above your threshold (300,000 weighted tokens by default) Claude does not start. Quotient works out the numbers of each level itself, so none is left out, and you can write a choice of your own in the free field.
Each level says how long it should take and what it would leave of your limits. In the words of the first user, during two installments: "I know how long it takes, so I can step away from the computer."
Claude tells you at your next message. The cost comes from Claude Code's own records, and the time counts only the work, without your turns.
It learns only from clean jobs: chat jobs where you wrote nothing while Claude worked. Each type of work has its own factor, and until a type has 5 clean jobs it starts from the number in the picture.
A notification and a report in the job's chat tell you what each piece did. Several jobs together never eat the share of the week you keep for yourself. Waking and sleeping work on Windows; on macOS and Linux Quotient gives you the line to schedule it yourself.
It can sit in Claude Code's status line (setup-statusline --write), and it works in the desktop app too. Anthropic does not publish the limits in tokens, so the token numbers are an estimate, measured on your PC or, until then, the shared size of your plan.
At every step Claude re-reads everything the chat holds. Quotient measures that apart from the work, and when a chat has grown long the window offers a new chat and says how much that saves.
Turn on config tips.level 1 or 2 and Quotient also tells you where a chat's tokens went, tool by tool, and what would save most, in that order. You can put a tip off or drop it, and no tip ever asks for shorter replies.
Copies of Quotient pool one line of numbers per finished job. The average fills in where you have no numbers yet: your first installments and the size of your plan's limits. Chat quotes start from the measured factors above. Since 1.2.3 it can also stay a small part of your factor (shared_min_weight), off by default.
Hermes Agent is another AI agent, which can run on the same Claude login. Two hooks in its config are enough, and only Claude chats count against the plan. The setup is under More details.
Install
In Claude Code:
/plugin marketplace add Korvonordico/quotient-claude-tokens
/plugin install token-quotient@token-quotient
You need Python 3.8 or newer (standard library only, nothing else to install) and sh, which on Windows comes with Git for Windows, already used by Claude Code.
With OpenAI Codex (experimental)
Token Quotient also works inside Codex, OpenAI's coding agent (the codex program in a terminal). Since 1.3.7 it is a Codex plugin, and Codex itself does the installments. It is new: it has been tested on sample session logs and with a stand-in for Codex, not yet in a long real use. Please open an issue when something looks wrong.
Install it as a plugin (the way to choose):
- In a terminal:
codex plugin marketplace add Korvonordico/quotient-claude-tokens. This adds Token Quotient to the plugin lists Codex knows.
- Open Codex, type
/plugins, pick Token Quotient and install it. Then start a new Codex session.
- Type
/hooks and trust Token Quotient's three hooks. Codex skips new hooks until you trust them, and without them nothing is measured.
- Ask Codex: "set Token Quotient up". It asks for the threshold, the weekly reserve and the language.
以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗
