Prompts

A prompt is a reusable prompt template the server defines and the client instantiates. Where a tool does something and a resource is something to read, a prompt is guidance — instructions the server maintains so every client doesn't have to re-implement them.

Instantiating a prompt returns a list of messages you prepend to your model's context.

Discover prompts

const { prompts } = await client.listPrompts();
// → [{ name: "...", description: "...", arguments: [...] }, ...]

Each prompt exposes:

• name — the stable identifier, used in prompts/get
• description — when and why to use it
• arguments — the parameters it accepts (name, description, whether required)

Get a prompt

const result = await client.getPrompt({
  name: "<prompt-name>",
  arguments: {},
});
// result.messages — role/content pairs to prepend to your conversation

The server fills the template with your arguments and returns the resulting messages. Prepend them to your own message history before you call your model.

Prompt, tool, or resource

Use case	Pick
Agent behaviour — routing, stop conditions, source rules	prompt
Searches, lookups, anything parameterized	tool
Reference data the agent reads on demand	resource

A prompt is instructions (how to act), a tool is a verb (do something), and a resource is a noun (read something).

When to reach for a prompt

Server-defined prompts shine when the server knows best how its own tools should be used:

• Shared policy. Routing rules, stop conditions, and source-attribution guidance live in one place instead of being copied into every agent.
• Updatable without redeploys. Change the prompt server-side; clients pick up the new guidance on their next listPrompts call.
• Capability-coupled. The server can encode how to compose its own tools — knowledge that would otherwise be hardcoded in each client.

What EDITED MCP serves

The prompt catalogue is environment-specific and grows over time, so rather than hardcode it here: discover it at runtime with listPrompts (above), or browse the Reference, which is generated from the live server and lists exactly what your gateway exposes.