MCP Server Capabilities

Goal

Leverage advanced MCP features to provide agents with direct data access, reusable prompt templates, and the ability to request completions or user input from the host.

1. Resources (Contextual Data)

• •Definition: A server-side capability allowing the host application to read static data (logs, file contents, database records) directly.
• •Usage: Use when the agent needs to "read" the state of the world rather than "act" on it.
• •Mechanism:
  • •URI-Based: Resources are identified by URIs (e.g., file:///logs/error.txt).
  • •Embedded vs. Linked: Resources can be returned fully embedded in the tool result or as a link to be fetched separately.
• •Security Note: Always validate resource URIs against an allowlist to prevent arbitrary file access.

2. Prompts (Reusable Templates)

• •Definition: Pre-defined prompt templates stored on the server that clients can retrieve and use.
• •Usage: Use to standardize how agents interact with your tools (e.g., a "Bug Report" prompt that pre-fills the necessary context for the create_ticket tool).
• •Risk: High risk of prompt injection. Clients should treat server-provided prompts as untrusted user input.

3. Sampling (Server-Initiated Intelligence)

• •Definition: A capability that allows the Server to ask the Client (Host) to run an LLM completion.
• •Workflow:
  1. •Tool needs complex reasoning (e.g., summarizing a massive log file it just fetched).
  2. •Server sends a sampling/createMessage request to the Client.
  3. •Client runs the model (potentially with human approval) and returns the text.
• •Benefit: Offloads compute costs to the client and keeps API keys secure on the client side.

4. Elicitation (User Input)

• •Definition: A mechanism for the Server to ask the Client to prompt the human user for input.
• •Usage: Use for disambiguation (e.g., "Which project did you mean?") or required parameters that are missing.
• •Privacy: Servers MUST NOT use elicitation to request sensitive information (passwords, API keys).