Concept · Reeve
Kept off the cloud
How Reeve counts the tokens it read on this PC so your AI coding assistant didn't have to, what each tool counts, and why the numbers are conservative.
- Article
- 1208
- Applies to
- Reeve 0.17.2
- Last reviewed
- For
- For developers
What it shows#
Every token Reeve reads on this PC is one your AI assistant didn't have to: it got a short answer instead of the whole log, the files or the docs. Reeve's page shows the total as Kept off the cloud: this month, last month and all time, about how many tokens and about what they'd have cost, with what each tool read.
Your assistant's status tool ends with the same totals in one line.
The price is what your assistant's model costs per million input tokens: $3.00 unless you set another in Reeve's settings. On a plan rather than paying per token, you save usage limits rather than money: the tokens are the same.
How it counts#
The numbers are estimates, kept low on purpose, and always shown as "about":
- A token is 4 characters, rounded up, for everything. That's fewer tokens than models count.
- Per call, Reeve counts what it read on your assistant's behalf and what it sent back. What's kept is the difference, never below zero, so a call that answered at length takes nothing off another's.
- Only what came from this PC's disk counts. Text your assistant passed in the call itself was already in its context, so it saves nothing.
- Caps: one input (a file, a log) counts for at most 25,000 tokens, and one call for at most 100,000. Faced with more, an assistant would search it rather than read it all.
- Failures and repeats: a call that fails counts nothing. The same call, reading the same amount, within an hour in the same session, counts once.
What each tool counts as read#
| Tool | Read |
|---|---|
triage | The log, as the model would be shown it, even when code alone settled the outcome |
summarize | The files |
map | The parts of each file the model was shown; nothing for a file it didn't reach |
docs | The doc passages the model read |
search | The code in the ranges it found: a stand-in for the looking it saved |
cluster | The items |
outline | The files it outlined |
changes | The diff |
ci | The failed jobs' logs |
refs | The search output it read |
history | The log it read |
pr | The pull request and its review threads |
layout | The package and README files it read |
session | The session's transcript, condensed, as the model is shown it |
Not counted: ask (your assistant writes its input); status, digest, repos, manor and docs_list (Reeve's own records); commands, errors and sessions (history an assistant wouldn't read whole); lint_docs and upgrades (code checks and release notes with no fair comparison); and screens (pictures).
What's stored#
Counts only, in kept.json in Reeve's data folder: per day and tool, the calls, and the tokens read, returned and kept. Never an input, an answer, a path or a prompt. Days older than 90 are rolled into their month, so the file stays small.
Related articles
Is this page right?
If something on it is wrong or out of date, tell us and we'll fix the page.
Still stuck? Write to support@castellan-software.com and mention article 1208. Every version of Reeve, and what changed in it, is in its release notes.