RSS
Research

ystack says its Claude Code mods cut usage 78%. Its own write-up calls that the upper end

ystack, a new bundle of 35 Claude Code mods, says it cut usage about 78% per hour of work. The figure comes from one heavy user's logs, and the author treats 39% per API call as the firm number.

Close-up of a car fuel gauge showing empty, illustrating ystack's claim of cutting Claude Code usage
Photo by Wesley Tingey on Unsplash

ystack is an open-source collection of mods, skills and setup tools for Claude Code, Codex, OpenCode and Antigravity, published on GitHub under the MIT license. Its README leads with a big number: on one heavy user's logs, it cut usage about 78% per hour of work and 39% per API call.

The project links to a long write-up of how that was measured, and it is more careful than the headline. The numbers are a before-and-after of one person's real usage while the rules and mods went in, not a controlled test.

What the logs show

  • About 248,000 API calls from September to November, main sessions and subagents combined.
  • Usage per active hour fell from 38.4M to 8.3M input-token equivalents, and median context per call halved from 375k to 189k tokens.
  • The before period covers 37 active hours, the earliest logs that survive. The after period, with the rules enforced by the mods from November 7, covers 63.
  • The hourly drop also reflects fewer agents running at once. The author says to treat 78% as the upper end and 39% per call as the firm floor.

Where the usage actually went

The more useful part is a separate breakdown of 14 days on Opus, about 60,000 API calls and 1.9B input-token equivalents. In that sample, subagents made 73% of usage. Six sessions that ran for 31 to 219 hours made 91% of it, and five of them sat near the point where compaction kicks in. The prompt text added by the roughly two dozen mods installed at the time was 0.3% of context tokens.

A replay of those 14 days that changed only one thing, handing off long sessions at 252k tokens and compacting subagents at 250k, cut usage from 1.90B to 1.61B input-token equivalents, about 15%. The author calls the main-session part optimistic, since it assumes a handoff every time a session crosses 252k. The replay also leaves out two earlier changes it can't model: building at medium effort instead of high, and turning agent teams off. So the before-and-after figure doesn't show how much of the drop each change made.

What to take from it

If your usage looks like this user's, the levers the write-up points at are not exotic: high-effort subagents made 59% of one day's usage, and agent teams' idle pings made 32% of one 9-day session. ystack's /ystack-stats command estimates what it would save by replaying your own last two weeks of logs, and its claude/tools/usage-report.py script reports what you actually used. Either is a better guide than someone else's 78%.

python claude/tools/usage-report.py --days 14

Installing the whole bundle adds 35 mods, and some act on their own from the start. Per the README, context-meter swaps a very long conversation for a short handoff note, keep-going sends Claude back to work when it offers to continue, and push-guard refuses risky pushes. The project calls itself an early 0.1 release. Read what each mod changes, and install only the ones you want.

Source: ychampion/ystack README, ystack: long-session measurements

This story first appeared in the claudefox daily briefing for November 14, 2026.