Your page was true the day you published it. Is it still?
A security note says a CVE is not being exploited. A study says its counts come from a dataset with this checksum.
A release page says the tarball is the one you signed. A comparison says the vendor offers a free tier. Each was
true when written, and the world kept moving. claim-watch rechecks every such claim against its outside source,
daily or on demand. One file, no dependencies, no keys, no model. It reports; it never edits a page.
$ node watch.mjs claims.json
claim-watch, 2026-11-07 13:05 UTC. 3 of 4 claims hold; 0 of 0 outside links answer.
NO LONGER HOLDS (1): the page needs a dated correction, written by a person:
- [demo] CVE-2021-44228 is not known to be exploited: listed on 2021-12-10 (catalogue 2026.11.06)
This report changes no page.
(A real run against CISA's live catalogue. The claim is Log4Shell, written to fail on purpose.)
| check | the claim | decided by |
|---|---|---|
served |
what a stranger receives is the file you reviewed | the live URL, byte for byte against a local file |
sha256 |
an outside document has not changed | its checksum; 404/410 = gone |
page_has |
a source still says it | strings on the page |
page_lacks |
a source still does not say it | a string absent, with must_have to catch a layout change |
json_equals |
an API still reports a value | a dotted path in the JSON |
github_asset |
a release file was not replaced | GitHub's sha256 digest for the asset |
kev_absent |
a CVE is not known to be exploited | CISA's Known Exploited Vulnerabilities catalogue |
link_pages |
every outside link still answers | 404/410 or a dead name = gone; preconnect hints are not links |
See claims.example.json. Leave sha256 / digest blank and run once with --record to fill them from today's
source. --record never overwrites a baseline, so it cannot hide a change.
git clone https://github.com/PowerpBoxCaptain/claim-watch ~/.claude/skills/claim-watch # as a Claude Code skill
node watch.mjs claims.json [--out report.txt] [--record]Daily from cron, with the report kept where your agent or you will read it:
0 7 * * * node ~/claim-watch/watch.mjs ~/site/claims.json --out ~/site/claim-watch.txtExit codes: 0 every claim holds · 1 a claim no longer holds, or a link is gone · 2 none failed, but a claim
could not be checked (not a pass) · 3 broke. A source that does not answer is never counted as holding.
It does not find the claims for you, prove they were right on day one, or edit the page. A page that needs fixing gets a dated correction from a person. Sites that block automated requests show up as "not reached" or "could not check", never as passing.
node test.mjs runs 28 cases against a local server whose answers change between runs: every check holding and
failing (a positive control for each), gone documents, server errors counted as not checked, --record filling a
blank baseline and refusing to overwrite one, a layout change caught by must_have, dead links and dead names,
preconnect hints ignored, an unreachable source, and a broken list.
Claude Code (this repository is a one-plugin marketplace):
claude plugin marketplace add PowerpBoxCaptain/claim-watch
claude plugin install claim-watch@claim-watchCursor (local plugin): git clone https://github.com/PowerpBoxCaptain/claim-watch ~/.cursor/plugins/local/claim-watch, then run
Developer: Reload Window.
GitHub Copilot: copy .github/agents/claim-watch.agent.md into the same folder in your repository.
What has been tested: in Claude Code, the plugin was loaded and given a real task in an isolated session, its tool
calls were checked against SKILL.needs.json, and the two commands above were run. The Cursor and Copilot files are checked against those
clients' documented formats only; they have not been run in Cursor or Copilot yet. The plugin carries the skill only:
it installs no hooks, so any hook described in this README is still set up by hand.
AGPL-3.0-only. See LICENSE.