Skip to main content

MCP Prompts

What are Prompts?​

Prompts in the Model Context Protocol are pre-built, reusable templates that MCP servers expose to help users interact with LLMs. They provide structured starting points for common tasks.

Specification: MCP Prompts Spec

How It Works​

Servers can expose templated prompts with:

  • A descriptive name and purpose
  • Optional arguments for customization
  • Pre-configured messages for the LLM

When you use a prompt, the server fills in the template and sends the formatted message to your LLM.

Configuration​

Prompts are discovered automatically from MCP servers:

mcpServers:
  code-helper:
    type: stdio
    command: npx
    args: ["-y", "my-code-mcp-server"]

If the server supports prompts, they'll be available immediately.

Using Prompts​

In Web UI​

Type / to discover and invoke prompts from connected MCP servers.

In CLI​

List available prompts:

dexto
> /prompts

Use a prompt:

> /use code-review file=src/app.ts

Or use the shorthand (if supported):

> /code-review src/app.ts

Via SDK​

// List prompts from a server
const client = agent.mcpManager.getClient('code-helper');
const prompts = await client.listPrompts();

// Get and execute a prompt
const prompt = await client.getPrompt('code-review', { file: 'app.ts' });
const response = await agent.sendMessage(prompt.messages);

Prompt Structure​

Prompts can include:

  • Text content - Instructions and context
  • Images - Visual references (base64-encoded)
  • Resources - Embedded file contents from the server

Arguments can be:

  • Required - Must be provided by the user
  • Optional - Have default values
  • Auto-completable - Server suggests valid values

See Also​