Skip to content

Latest commit

 

History

History
85 lines (50 loc) · 4 KB

capabilities.md

File metadata and controls

85 lines (50 loc) · 4 KB
title
Capabilities

Sampling

MCP servers can request LLM completions from connected clients that support the sampling capability. This lets your tools offload summarisation or generation to the client’s model.

For a runnable server that combines tools, logging and tasks, see:

In practice you will:

  • Declare the sampling capability on the client.
  • Call server.server.createMessage(...) from within a tool handler.
  • Return the model’s response as structured content and/or text.

Refer to the MCP spec’s sampling section for full request/response details.

Elicitation

Form elicitation

Form elicitation lets a tool ask the user for additional, non‑sensitive information via a schema‑driven form. The server sends a schema and message, and the client is responsible for collecting and returning the data.

Runnable example:

The simpleStreamableHttp server also includes a collect-user-info tool that demonstrates how to drive elicitation from a tool and handle the response.

URL elicitation

URL elicitation is designed for sensitive data and secure web‑based flows (e.g., collecting an API key, confirming a payment, or doing third‑party OAuth). Instead of returning form data, the server asks the client to open a URL and the rest of the flow happens in the browser.

Runnable example:

Key points:

  • Use mode: 'url' when calling server.server.elicitInput(...).
  • Implement a client‑side handler for ElicitRequestSchema that:
    • Shows the full URL and reason to the user.
    • Asks for explicit consent.
    • Opens the URL in the system browser.

Sensitive information must not be collected via form elicitation; always use URL elicitation or out‑of‑band flows for secrets.

Task-based execution (experimental)

Task-based execution enables “call-now, fetch-later” patterns for long-running operations. Instead of returning a result immediately, a tool creates a task that can be polled or resumed later.

The APIs live under the experimental .experimental.tasks namespace and may change without notice.

Server-side concepts

On the server you will:

  • Provide a TaskStore implementation that persists task metadata and results.
  • Enable the tasks capability when constructing the server.
  • Register tools with server.experimental.tasks.registerToolTask(...).

For a runnable example that uses the in-memory store shipped with the SDK, see:

Client-side usage

On the client, you use:

  • client.experimental.tasks.callToolStream(...) to start a tool call that may create a task and emit status updates over time.
  • client.getTask(...) and client.getTaskResult(...) to check status and fetch results after reconnecting.

The interactive client in:

includes commands to demonstrate calling tools that support tasks and handling their lifecycle.

See the MCP spec’s tasks section and the example server/client above for a full walkthrough of the task status lifecycle and TTL handling.