MCP registry drift feed
What a line on this page means — and what it does not
It means: the registry’s published entry for that server was different the second time we read it, in the way described. Every row links to the registry’s own API response for that server, so you can check us against the source rather than take our word for it.
It does not mean a tool’s schema changed. This crawler has never connected to an MCP server and has never seen a tool schema. The registry publishes packaging and transport metadata — that is all we read, and all we can speak to.
It does not mean anything is broken, unsafe, or wrong. Publishing a new version is the normal, healthy behaviour of a maintained server. Nothing on this page is a judgement about quality, security, or correctness, because nothing we observe could support one.
Coverage record unreadable
We could not read our own record. The store this page reads did not answer, so we cannot tell you what we have observed — or even whether we have ever looked. That is a gap in our own plumbing, not a fact about the registry, and it is not the same as an empty feed. Everything below is withheld rather than guessed at.
All observed changes
We could not read our own record, so this list is withheld — see Coverage above.
Crawl log
Published because a feed whose gaps are invisible cannot be read honestly. A quiet stretch above means one of two very different things, and only this table tells them apart: the registry was quiet, or we never looked.
We could not read the crawl log either, so this table is withheld rather than shown empty.
ok — the sweep reached the last page. partial — it stopped early (page budget, or an upstream error mid-sweep); what it read is kept, and the next sweep resumes where it stopped. failed — the first request failed, so the run recorded nothing at all. running — a sweep that never reported a result; left visible rather than quietly resolved into a status nobody observed. Entries skipped counts registry entries whose shape this crawler did not fully recognise; they are dropped rather than guessed at, and counted rather than hidden.
How this page can be wrong
- We only see what the registry publishes. A server whose tools change without its registry entry changing produces nothing here. Silence about a server is not evidence about that server.
- We record net differences, not every intermediate state. If an entry changes twice between two of our reads, you see one line describing the difference between the endpoints, not the path.
- Servers outside the official registry are not tracked at all. Plenty of widely-used MCP servers are never listed there.
- A server disappearing from a response is never read as deletion. Only an explicit deleted or deprecated status from the registry produces that line.
- The observation date is ours, not the publisher’s. It is when we saw the change, bounded above by our crawl schedule.
Why we run this
Reelier records what an agent run actually did and whether it stayed inside its declared scope. Tracking what the tools underneath a run are publishing is the same instinct pointed one layer down — and it is the same rule in both places: state what was observed, name what was not, and never let the artifact imply more than it proves.