Report · October 2026

The server you approved keeps shipping.

We launched up to the newest 12 stable releases of each of the 30 most-downloaded MCP servers, March to September 2026, and compared the tools each one offered your agent. 16 of the 30 start without a key; across their 107 updates, 49 (46%) changed them, and nothing asked you to look again.

30 servers185 versions launchedprobed 29 Sep – 22 Oct 2026run locally

49 of 107 updates changed the tools.

  1. Server A8 of 11Server A: 8 of 11 updates changed the tools.
  2. Server B7 of 11Server B: 7 of 11 updates changed the tools.
  3. Server C7 of 10Server C: 7 of 10 updates changed the tools.
  4. Server D6 of 10Server D: 6 of 10 updates changed the tools.
  5. Server E6 of 10Server E: 6 of 10 updates changed the tools.
  6. Server F4 of 9Server F: 4 of 9 updates changed the tools.
  7. Server G4 of 4Server G: 4 of 4 updates changed the tools.
  8. Server H3 of 11Server H: 3 of 11 updates changed the tools.
  9. Server I2 of 11Server I: 2 of 11 updates changed the tools.
  10. Server J1 of 2Server J: 1 of 2 updates changed the tools.
  11. Server K1 of 1Server K: 1 of 1 updates changed the tools.
  12. Server L0 of 9Server L: 0 of 9 updates changed the tools.
  13. Server M0 of 4Server M: 0 of 4 updates changed the tools.
  14. Server N0 of 2Server N: 0 of 2 updates changed the tools.
  15. Server O0 of 1Server O: 0 of 1 updates changed the tools.
  16. Server P0 of 1Server P: 0 of 1 updates changed the tools.
The 16 of the 30 that start without an API key. One square per update from one stable release to the next, 10 March to 9 September 2026, in release order. Changed means any tool's name, description, input schema or annotations differed; an update counts once however many tools moved. Six servers held 38 of the 49 changes; five never moved in 17 updates. Servers are not named; findings go to the vendors privately.

How it reaches you

You approve once. The next launch runs whatever shipped since.

29 of the 30 document an install line with a bare package name or @latest. Your client starts the server with that line, so the next launch after a release runs the new version. The 30th documents only a hosted URL, which updates on the vendor's side instead.

Each square is one of the 30 servers. Filled: its README's install line is unpinned. Dashed: hosted URL only.

"args": ["-y", "some-mcp-server@latest"]

Your agent reads the tool list the server offers on the day and calls it. The model takes each description as an instruction and each parameter as a capability. Nothing in that loop asks you a second time.

What the 49 changes did

Half of them changed what a tool is, does or tells the model.

  1. 4adds a parameter that names a new destination or runs something
  2. 2changes a tool's safety label
  3. 7adds or renames a tool
  4. 12rewrites a tool's description
  5. 23changes parameters, adds nothing
  6. 1only removes tools
Each update counted once, by its most serious change. Darker: changes to what a tool is, does or says to the model (25). Lighter: edits inside existing parameters, or removals (24). One of the 12 rewrites was whitespace only.

Counted at parameter level across all 49: 94 descriptions rewritten, 77 parameters added, 16 removed, 5 types changed. A rewritten description is a different instruction. A new parameter is a different capability. Your agent takes both at face value.

Retired, still installed

Six of the 30 were deprecated. People kept installing them.

Downloads of the six deprecated packages, 21–27 Sep 2026455,402
…of which the one with a published SQL-injection finding109,622
Downloads include CI and mirrors; they are not users. npm's daily series has gaps, so the week's total is a floor. Deprecation checked on 22 October.

Five of the six READMEs still showed npx -y; the sixth's install line was unpinned too. For the one with the SQL-injection finding, the fix was merged the day its repository was archived and never published: the version npm serves predates it.

Labels you cannot trust

A tool's safety label is a hint. We found it wrong both ways.

Annotations are the only hint a client gets about what a tool does. We read them across the 16 keyless servers on 22 October.

The label said
destructive
What it was
Two read-only searches. The server derives the flag from the HTTP verb, and those searches are POSTs.
The label said
read-only
What it did
Overwrote a collection, when called from a client that does not declare elicitation support. From one that does, the server asked first, and declining left the data intact.

The label was wrong either way; the client decided what happened. The overwrite ran against a throwaway local database on this machine. No third-party data was touched.

Cost as a side effect

Same 27 tools. Nearly three times the context on every prompt.

Version 2.x, context cost of its tool list5,003 tokens
Version 3.x, same 27 tool names14,232 tokens
Counted with cl100k: a comparable ruler, not exact billing for any model.

Nothing new to call. Two of its tools grew from about 600 tokens each to over 2,800, and the bill for that arrives on every prompt, whether or not the agent uses them.

Method and caveats

How we counted, and what this does not cover.

The 30 most-downloaded official MCP servers on npm, by weekly downloads for 21–27 September 2026. Every stable release from 10 March to 9 September; only the newest 12 stable versions per package were launched.

Each version ran in a throwaway home directory with no API keys in the environment and answered initialize and tools/list only. Nothing else was executed, except the overwrite test above.

Launching without a key understates servers that gate tools on one: one server showed 25 tools keyless and 27 keyed. Seven of the 30 would not start without a key, and seven had no stable release in the window; they contribute nothing to the 107.

An independent study of nine servers over the same window measured 48%. Our like-for-like re-run of the same nine gave 37 of 77, also 48%, with eight of nine servers matching her per-server counts exactly.

Data was recorded between 29 September and 22 October 2026 and is true as of those dates. Servers are not named here; per-vendor findings go to the vendors privately.

Record your baseline.

The free CLI reads the configs your agents already have and pins what each server offers today. When a tool changes, the call waits for you. Your inventory stays on your machine.

uv tool install --force mcpgawk
mcpgawk

Hold the change before it runs.

The gateway sits in the call path. A tool that differs from the one you approved is held until a person says yes.

Start the 7-day trial →

No card. Per person.

Written up from this data: I re-ran 152 versions of popular MCP servers · Half the updates changed the tools. Which ones should stop your agent.

Probed locally with mcpgawk's pin logic and the research scripts in its repository. Free and open-source · GitHub · gawk.dev