Reference
Extra checks
Three results that come with every connector check but sit outside the twelve steps: which MCP versions your server accepts, what changed in your tools, and what your tool descriptions say to the AI.
#Other MCP clients (protocol versions)
Claude is not the only MCP client, and clients ask for different protocol versions. A server that only speaks a newer version can work in one app and fail in another. MCPatch sends the initialize request once for each published version and reads what comes back:
2024-11-05 2025-03-26 2025-06-18 2025-11-25| Result | When |
|---|---|
| pass | The server answers with the same version the client asked for. |
| warn | The server answers with an older version. That works only if the client also supports the older one. |
| fail | The server answers with a version the client does not know, an error, or nothing usable. |
| skip | Skipped when the handshake failed, or when the server needs sign-in and no token was given. |
The overall result is a pass when all four versions are accepted and a warning when any is not. The fix: let your MCP SDK negotiate, answering initialize with the version the client asked for when you support it. If you upgraded an SDK that dropped older versions, pin the previous release.
#Tool changes
A client that already loaded your tool list keeps calling the old shape. If you rename a tool or change a parameter, the handshake still passes but calls start failing. MCPatch saves a compact snapshot of your tools on every check and compares it with the previous check of the same connector (your browser sends the id of its last check; if it does not, MCPatch falls back to the last check from the same visitor).
What counts as a breaking change
- A tool was removed.
- A parameter was removed from a tool.
- A parameter changed type.
- A parameter became required.
What counts as safe
- A tool was added.
- A new optional parameter was added.
- A parameter became optional.
- A description changed.
Breaking changes give a warning with the list of changes. Only safe changes give a pass. On your first check MCPatch just saves the snapshot and tells you to check again after your next deploy. The snapshot keeps tool names, parameter names and types, required parameters and the first 160 characters of each description, never your data.
#Hidden instructions in tool descriptions
An AI reads every tool description and parameter description as if it were an instruction. A malicious or careless server can hide text there that steers the AI, a tactic known as tool poisoning. MCPatch scans each description for text aimed at the AI, such as:
- hidden instruction tags;
- text that tries to override your agent's own instructions, or tells the AI to hide things from you or act without telling you;
- orders about how to use other tools;
- invisible characters: Unicode tag characters and text-direction tricks count as serious, zero-width characters as medium;
- references to throwaway data-collection hosts or to private keys.
Any serious finding gives a fail, any other finding a warn, and no findings a pass. Each finding names the tool, where the text is, why it matters and a short excerpt.