State of the Agent Economy — Q3 2026

Nearly 1 in 3 MCP servers — the tools AI agents call to act in the world — is already dead or no longer maintained. One in five is fully dead: the repository deleted or abandoned outright. Not an estimate — a full census of 28,824 public MCP-server repositories, re-checked daily on a live panel.

Is your stack affected?

Check your own dependencies against the register — no signup, no key:

npx lastseen-cli check

reads your project's package.json and dates every dependency. Or ask about any component directly:

curl https://lastseen.dev/api/v1/entity/{owner}/{repo}

Search a specific server from the front page. A question and a tool — we report dated facts; we don't tell you what to do about them.

Three numbers from the census

1,027 servers changed a tool's schema under a stable name. Between two dated reads, 1,027 MCP servers altered a tool's declared inputs or description while keeping the same name — the observable substrate of a schema-swap. We date the change; we do not judge intent.

98% still speak an older protocol. Of the servers probed for protocol version, only 2.2% have adopted the current (2026-07-28) MCP spec. Fewer than 1 in 40 is on the latest generation.

Of the thousands of servers observed dead, only 6 ever came back. Across the panel, 3,510 MCP servers have been observed deleted; exactly 6 later returned to life — about 1 in 585. And when a dead name does return, it is often a different project reusing a vacated slot. In this economy, death is very nearly permanent.

What the full report covers

Mortality, correctly defined — 12% of servers are deleted/inaccessible and 1 in 5 is fully dead; “abandoned” headlines measure staleness, death is a different, smaller number. Both, on explicit denominators.
Migration Watch — a dated baseline of protocol adoption: the first fixed point from which real migration can be measured, not a rate claimed today.
Mutations & resurrection — 998 dated state changes; 6 servers returned from the dead (a reused namespace is a supply-chain discontinuity a snapshot hides).
The substrate — 4.7M models, spaces, and datasets under the agent layer, linked by 2.87M dependency edges.

Read the full report → · Dataset (CC-BY-4.0): Zenodo · Verify any dated observation: /verify

Citeable numbers

Each line is a complete, sourced sentence.

“Nearly 1 in 3 public MCP servers is dead or unmaintained, and 1 in 5 is fully dead — a census of 28,824 repositories (lastseen.dev, Q3 2026).”
“1,027 MCP servers changed a tool's schema under a stable name (lastseen.dev, Q3 2026).”
“Only 2.2% of MCP servers have adopted the 2026-07-28 protocol; 98% still speak an older generation (lastseen.dev).”
“Of 3,510 MCP servers observed dead, only 6 ever returned — about 1 in 585 (lastseen.dev, Q3 2026).”
“Every figure is re-derivable point-in-time: append ?as_of=YYYY-MM-DD to any entity URL.”

Method, denominators, caveats, and hash-chain verification are in the full report. Corrections: /dispute.