Shipped this week: Limitcast 0.18.1 is available on npm and GitHub. It makes trimming more conservative: trim --apply now disables only MCP servers that have been quiet for 30 days. A server unused during the report window but used within those 30 days stays enabled. An explicitly empty keep-list requires --yes before applying changes.
The release also makes results lead with the net change in the starting context you control, and fixes the Codex guard's lookback around local midnight. These changes are in the published package; main currently has no additional source-only changes beyond the release.
Also shipped since last week's update: 0.18.0 added duplicate-subagent holds, optional tool-result limits, and the trim --share privacy fixes previously available only on main.
Actual public npm-registry output checked for this update on November 6, 2026; no synthetic usage figures or private session data:
$ npm view limitcast version
0.18.1
limitcast results --days 7
Use it for a weekly review of starting-context changes, forecast evidence, and recorded protection events. Install or update with npm install -g limitcast on macOS or Linux with Node.js 20+. For Claude usage readings, run limitcast install and restart Claude Code; useful comparisons need recorded history, so a new installation may have little to report.
Context estimates are not measured cash savings or proof that Limitcast caused a reduction. Recorded holds are events, not saved-token totals. A configuration check alone does not prove a live guard hold.
Does the 30-day inactivity threshold keep the MCP servers you still need enabled, or is it too conservative? Tell us which coding tool you use and the cutoff you would find useful, without sharing private logs.
Follow shipped releases with Watch → Custom → Releases and enable email delivery in your GitHub notification settings if you want release emails.