What do the feature flag values mean?
Less than they look like. A value here is what Anthropic's flag server returned for one account, on one version, at the moment this site read it. It cannot tell you whether a feature is on for you.
Claude Code evaluates flags remotely: the client sends its attributes, the server decides, and only the answers come back. The targeting rules and rollout percentages never reach the client at all. So a flag reading false here can be at 95% rollout globally with this account in the holdout, and no client can tell the difference. Reading a value here as "the feature is off" is the mistake this page exists to prevent.
The account behind every reading here is a 20x Max account authenticated with a claude setup-token long-lived token, which isn't what a typical install looks like and may be targeted differently. The attributes it sends include subscriptionType, rateLimitTier, organizationRole, email and releaseChannel. Any of them can move a value.
That skew has a size. Against an anonymous baseline from the same endpoint, 47 of 492 flags come back different for this account, in both directions: tengu_flint_harbor is true anonymously and false here. So it's a different sample, and being logged in doesn't make it a truer one.
One more trap. A flag missing from the payload isn't off. The code falls back to whatever default was compiled into it, which is sometimes true. Of the 341 gates the v2.1.232 bundle reads, 116 have no key in the payload at all. This site labels "no value" and "disabled" differently, because they're different.
Is a changelog here complete?
No. Each release is read by model agents, and those runs aren't repeatable: the same prompt over the same release came back with 57 turns one time and 136 another, and reproduced 4 or 5 of its own 9 findings while turning up 3 to 5 new ones. Every changelog here is missing things a second pass would have caught, and which things is luck. What's here was really found in the code. What's missing might just have been missed.
Old changelogs don't get rebuilt to fix that. A rebuild re-runs the whole read at real cost, so a page reflects the run that made it.
Where do the value and watch scores come from?
A model reads the finished entries and scores them, in a pass of its own after the writing is done. They're a reading, and two runs won't agree exactly. Use them to sort a long page. I wouldn't quote one as a rating.
Coverage is thin too. Scoring only started at v2.1.219, so most releases here have no scores, and an unscored entry has no score at all, which isn't the same thing as a zero.
Why does a release say Anthropic hasn't published notes?
The Official Changelog section on a release page is in one of three states, and it says which.
- Available: the version has its own entry in Anthropic's
CHANGELOG.md, shown word for word and linked back. - Pending: the version is newer than anything in that file, so their notes aren't out yet. This one fixes itself: once a newer entry shows up without it, the page flips to gap.
- Gap: the version is inside the range Anthropic has published and was never listed on its own, which is what a skipped prerelease looks like.
None of the three says anything about the release itself. A version with no official note is often just a quiet one.
Why can I vote on an entry but not see the totals?
Because there aren't many votes yet, and a visible count of two biases whoever reads next. The votes are recorded and they do reach me. Once the numbers mean something, they go up.
Do the documentation dates say when Anthropic changed a page?
No. They say when this site noticed. A poller reads the published docs on a schedule and records what moved since its last read, so a change carries the date of the poll that caught it. The edit itself happened somewhere between that read and the one before it, and longer ago than that if the poller was down.
Is this site affiliated with Anthropic?
No. It's unofficial and unaffiliated, and nobody there has reviewed any of it. Anthropic's own notes are reproduced word for word and marked as theirs. Everything else is my reading.
How much of an entry is a machine's opinion?
The wording and the judgement are mine. What sits under them is Anthropic's own shipped code, and every entry carries the strings it was drawn from, so you can check one against the source instead of taking my word for it. If it doesn't hold up, vote it down and say why.
Why do the counts for older releases look low?
An older pipeline parsed the bundle a different way, and those releases were never re-read by the current one. Their counts run low and don't compare cleanly with recent releases, so a trend line crossing that boundary is measuring my tooling rather than Anthropic's output.