About lastseen.dev

The machine economy changes silently — and agents break when it does. We keep the dated record: what was declared, and when it changed. So you can see the change, and act before it breaks something you depend on.

The problem

Agents are built against a server's declared list of tools. There is no versioning duty and no changelog requirement for these declarations — a server can change what it declares at any time, without notice. Calls to a tool that is no longer declared return an error, often with nothing surfacing it, in systems that run unattended. And without a dated register there is no way to establish afterwards what was declared, or when it changed. That is why this record exists.

What we are

An independent temporal notary for the machine economy. We take dated snapshots of publicly declared, machine-readable material — MCP servers, npm packages, agent manifests — every day, and preserve what was declared, and when.

Nothing here is an opinion. A record says: on this date, this is what the source publicly declared. On a later date, this is what it declared instead. That is the whole service.

The measured scale

The machine economy changes quietly. Tools vanish from a server's declared list, contracts change shape, packages stop resolving — and the agents that depend on them stop working with no notice and no record of what was there before.

8.56 % of the MCP servers we track stop resolving every month — measured, not estimated. Source: our own dated observation series, 28,824 observations over the 43 days from 22 July to 3 September 2026, gaps handled as intervals rather than assumed. That rate is roughly three to five times higher than the figure the field had been assuming.

And contracts change even when servers stay up: across the ~16,000 MCP servers we track, about 12 % change their declared tool contract in a given month. Measured 2026-09-15 over the window 2026-08-13–09-15, counting only changes observed at the same endpoint on both dates — comparing across endpoints would measure which address we asked, not what the server declared.

Someone has to keep the dated record. Retroactivity cannot be reconstructed after the fact: whoever wants to know what a server declared last month has to have looked last month. We looked.

What we never do

• We record what changed and when — never why, and never who is to blame.
• We publish our method, our coverage limits, and our corrections log.
• We do not rank, recommend, score fitness, or predict.
• Everything is checkable at /verify. Free, no accounts, no paywall.

Independence

We have no vendor affiliation, no sponsorship, and no commercial relationship with any entity we observe. The observation history is deposited outside our own control, with citable DOIs, so the record cannot be quietly revised later — not even by us: Zenodo archive.

Personal data

We process publicly declared, machine-readable material. Some identifiers in public registries contain personal names — for example a server published as io.github.<name>/<project>. Our legal basis is legitimate interest: documenting and archiving the public machine ecosystem so that changes in it remain checkable afterwards.

You can object to, or ask us to correct, any record through /dispute. Submissions are logged, answered, and linked from the record they concern. We do not profile, we run no accounts, and we do not sell personal data.

Erasure and pseudonymisation

If you object to a record on personal-data grounds and we confirm the objection, we pseudonymise the subject in that record: the dated observation stays, the personal name is masked (for example, an MCP server in the io.github namespace). We do not delete the observation, because the history would then say something untrue about what we saw. The name was never the point — the date was.

In this first series we publish records for organisations and products only. Where a server is published under a personal account name, we hold the observation and do not publish it.

Corrections

If a record is wrong, or if you operate something we have observed and want to add context, use /dispute. Submissions are logged and linked from the record. Where we can confirm you control the source, your text is published next to the record within 72 hours, unedited by us, dated and linked.

These are factual observation records derived from dated snapshots of publicly declared, machine-readable material. They contain no assessment of fault, intent, or fitness for purpose. Corrections and context: /dispute.