One of my MCP tools returned full database rows. After a few calls, the conversation was mostly JSON and the model started losing track of the question.
Return a short summary from the tool and expose the full record as a resource. The model sees the summary immediately. If the host supports MCP resources and chooses to expose one to the model, the client can resolve the URI when the task needs more detail.
{
"summary": "3 open incidents, highest severity: high",
"resource": "incidents://open"
}The model's context grows by one line per call instead of hundreds, while the full data remains available through a separate read. That second read still depends on client support, authorization and the resource remaining available.
Key takeaway
Tools should answer the question; resources can hold the details.
- MCP resources specification — resources are application-driven context exposed by servers and selected by host applications.