Prompt Intents in MCP

🧠 What are prompts in the context of MCP?

Prompts in MCP are declarative mappings between natural language triggers and agent-exposed actions.
They allow agents to know what they can be asked — and what each question is expected to activate.

A prompt block is not just an example:
→ it’s a signal that says: “I want agents to respond to this intent.”

🧩 Where are prompts declared?

In any .llmfeed.json, typically inside:

{
  "prompts": [
    {
      "intent": "sign-document",
      "keywords": ["sign this", "please sign", "legal signature"],
      "description": "Trigger digital signature workflow for authenticated user"
    }
  ]
}

✅ Why prompts matter

Feature Impact
intent Canonical action identifier (agent logic)
keywords Allows agents to recognize varied phrasing
description For user-facing explanation / UI generation

Prompts are central for disambiguation, action safety, and self-documentation.

🔐 Security model

Prompts do not execute.
They are mapped to capabilities (or UI) — and depend on the trust of the feed:

  • A prompt in an unsigned feed = suggest only
  • A prompt in a signed + scoped + certified feed = fully executable

🧠 Prompt matching flow (agent side)

  1. User input triggers natural-language detection
  2. Agent compares it to known keywords from loaded feeds
  3. If match found and trust is sufficient:
    • Agent maps to declared intent
    • Verifies if capability exists and is permitted
    • Proceeds or requests user confirmation

🌍 Good practice

  • Always provide intent, even if not callable
  • Use keywords in multiple languages if needed
  • Pair with capabilities[] in same feed
  • Group in mcp-api if access is gated

🔗 Related feed types

  • export.llmfeed.json → declares offline promptable actions
  • mcp-api.llmfeed.json → scoped prompts per key
  • mobile-app.llmfeed.json → vocal prompts that replicate app behavior

🧩 Want more?