Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

140 silent MCP contract changes, and how to pin them

Written November 5, 2026

An independent maintainer found 140 silent contract changes in MCP reference servers. What changed, and why pinning versions is the fix.

What the maintainer found

An independent maintainer pinned the tool contracts of the MCP reference servers — a SHA-256 digest over each tool's name, description, input schema, and icons, with the output schema pinned as a separate hash — and diffed the published releases. The issue, filed October 29, 2026, reports 140 silent contract changes across 66 version pairs of the @modelcontextprotocol/server-* packages, none of them announced in the respective changelogs. The count is the maintainer's own measurement, and this page reports it as such.

Two servers carry most of the weight. The filesystem server, server-filesystem 0.2.0 to 1.0.0, had 14 findings — 8 breaking, 5 new, 1 notation change — and the maintainer notes it is the most-installed server scanned. The server-everything test server, 0.1.0 to 1.0.0, had 40 findings: 23 new tools, 4 tools gone, the echo tool loosened, and churn across prompts and resources. The full report, with verbatim quotes, per-pair tables, and one-command reproductions, is published alongside the maintainer's rugsnare tool, and each finding can be reproduced with npx rugsnare history <package>, which starts every published version locally and diffs the captured contracts.

The reference servers matter beyond themselves because they are the template the ecosystem copies. A tool contract that changes without notice in a reference server teaches every server built from that template that silent change is acceptable practice — and an agent built against a server's tools can then break silently when the server updates under it, with no error pointing at the changed contract.

How to pin them

  1. Pin the exact version of every MCP server you depend on. Do not install a server by a floating tag or a latest pointer. Record the version that your agent was built and tested against, and treat an update as a change to review, not a background event.
  2. Read the changelog on every update — and expect it to be silent. None of the 140 changes in the maintainer's scan were announced in a changelog. When you update a server, diff the tool list yourself: compare the tools, their descriptions, and their input schemas before and after, rather than trusting the release notes to mention a change.
  3. Test tool calls after an upgrade. Run the calls your agent actually makes against the new version and confirm each one still accepts the same arguments and returns the same shape. A contract change that passes a smoke test of "the server starts" can still break a call your agent makes every session.
  4. Put a contract check in CI. The maintainer's suggestion is a short gate — a ten-line check that surfaces contract changes in the pull request itself, before the update lands. Whatever form it takes, the point is the same: a change to a tool's name, description, or schema should be visible in review, not discovered in production.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: A malicious MCP server probed cloud credentials · Catch bad agent behaviour chains with Quarry