Elicitation, sampling and roots

Some MCP servers ask the client for things while a tool runs. TestPion's MCP inspector supports all three client features, so these tools can be inspected and tested too.

Feature	The server asks…	In the inspector	In a test
Elicitation	the user for input, with a JSON schema (e.g. a date to book a visit)	a form built from the schema, with Accept, Decline and Cancel	elicitation: with the answer
Sampling	the client for an LLM completion (messages, a system prompt, a token limit)	the request, and a reply you write or draft with Draft with AI (the AI assistant's model); nothing is sent to a model until you choose to	sampling: with a fixed reply
Roots	which folders it may work in	the workspace folder	roots: with paths or file:// URIs

The inspector declares these capabilities when it connects. The server waits while you answer (at most 10 minutes; disconnecting cancels). Only give a server what you'd give it: the dialog says which server asks.

In tests

A type: mcp test answers with fixed values, so runs are repeatable. The capabilities are declared only for tests that set them.

name: Booking asks for the date
type: mcp
server: clinic-mcp
tool: book_visit
elicitation:
  action: accept            # accept (default), decline or cancel
  content: { date: "{{visitDate}}", urgent: false }
assertions:
  - { type: equals, path: $.date, expected: "{{visitDate}}" }

name: Summary uses the client's model
type: mcp
server: clinic-mcp
tool: summarize_notes
sampling: "Rex has a sprained left hind leg."     # or { text: …, model: … }
roots: ["./records"]
assertions:
  - { type: contains, expected: sprained }

What the server asked is recorded in the result (serverRequests) and in the test's trace. Without elicitation:, a server that checks the client's capabilities doesn't elicit; if it asks anyway, the request is declined.