Usage Guidance the Model Actually Reads
August 5, 2026 · InfraScout Team
Three changes to administering external MCP servers, all of them about the screen telling you the truth.
A usage guide, separate from the description
Many MCP servers publish their own instructions during the protocol handshake, and the AI reads those on demand. Plenty of servers publish nothing, and for those the registration's Description field looked like the place to explain what the server is for.
It was not. Description feeds the portal card and the tool-group summary — never the conversation. Careful guidance written there was never reaching the model.
There is now a usage guide field that does. Write what the server is for, which tools to reach for first, and any convention its arguments follow, and the AI reads it the way it reads a server's own published instructions. The two are independent: your guide survives a server that publishes nothing, and rediscovering a catalog never overwrites your text. A server card shows which sources it has.
If you have been maintaining a long description hoping the model would use it, move that text into the usage guide.
The allowlist asks what to withhold
A server with every tool permitted used to paint every checkbox empty, because "no restriction" is stored as an empty list. Withholding one tool from a twenty-five-tool server meant ticking the other twenty-four by hand.
The column is inverted. An unrestricted server now paints as all-ticked, and you untick what it may not expose.
Two guards come with it, because "untick everything" is not the opposite of "tick everything" here — one of them means allow all. Unticking every row is refused, and points you at what you actually want, which is disabling the server. And opening the picker on a server with an exhaustive explicit list and pressing Save no longer converts "these exact twenty-five" into "plus anything added later"; an untouched save round-trips whatever was stored.
Custom groups that reported themselves empty
A custom tool group can hold InfraScout's own tools, references to external MCP tools, or both. Three places only looked at the first kind.
A custom group whose members were all external tools showed 0 tools on its card. Its token estimate failed outright, and a mixed group's estimate silently counted only the InfraScout half. Worst of the three: running Estimate All Tokens against a mixed group deleted its external references as a side effect.
All three now resolve a group's external half the same way chat does, so the card count, the estimate, and the write-back reflect what a conversation actually receives.
Check your mixed custom groups
If you ran Estimate All Tokens against a custom group holding external MCP tool references, those references were deleted. The deletion was a write, so there is nothing to restore automatically — re-add them from the group's edit view. Groups you never ran the estimate against are unaffected.
Questions or feedback? Reach us at info@infrascout.cloud.