Feed Type: mcp.llmfeed.json

Purpose

This feed acts as the main capsule that an agent reads to understand:

  • what this site or API can do,
  • what prompts it should react to,
  • what trust scope and signature applies,
  • and how to interpret the declared behaviour.

Location

Recommended path:

.well-known/mcp.llmfeed.json

Canonical Structure

{
  "feed_type": "mcp",
  "metadata": {
    "origin": "https://example.org",
    "title": "Main MCP Capsule",
    "description": "This service exposes actions for document conversion",
    "generated_at": "2025-05-19T10:00:00Z"
  },
  "prompts": [
    {
      "intent": "convert_pdf",
      "keywords": ["convert my PDF", "transform PDF to text"],
      "description": "Triggered when user wants to extract text from a PDF file"
    }
  ],
  "capabilities": [
    {
      "name": "convertPdfToText",
      "method": "POST",
      "path": "/convert",
      "description": "Convert a PDF to plain text"
    }
  ],
  "agent_behavior": {
    "mode": "suggest-only",
    "requires_user_confirmation": true
  },
  "trust": {
    "scope": "certified",
    "certifier": "https://llmca.org",
    "signed_blocks": ["feed_type", "metadata", "prompts", "trust"]
  }
}

Fields

Field Required Description
feed_type ✅ yes Always "mcp"
metadata ✅ yes Basic info: origin, title, date
prompts optional Declared intents with keyword triggers
capabilities optional Declared APIs or actions callable by agents
trust optional but recommended Certifier, scope, signed_blocks
agent_behavior optional Declares expected UX

How agents use it

  • Recognize actionable prompts from user messages
  • Evaluate trust before executing anything
  • Call declared endpoints (via capabilities block)
  • Adjust their UX flow based on trust + agent_behavior
  • Use as handshake capsule for the site

Variants: mcp-lite.llmfeed.json

While the standard MCP feed is full-featured, sites may also expose a lightweight capsule under:

.well-known/mcp-lite.llmfeed.json

This file must still declare:

"feed_type": "mcp"

But is optimized for:

  • 💡 voice assistants,
  • 💾 low-bandwidth environments,
  • 🎯 agent previews or minimal understanding.

Recommended fields:

  • metadata
  • prompts[]
  • optional agent_behavior
  • (trust and capabilities often omitted)

Agents that find this file should interpret it as a lightweight fallback or teaser, not a full declaration of trust or functionality.

Optional: Agent Interaction Blocks

The following blocks may be declared inside mcp.llmfeed.json to help agents interact with a service or user support flow:

📮 agent_services

Define how users can trigger support or assistance requests.

"agent_services": {
  "keywords": ["support", "callback", "help me"],
  "action_endpoint": "https://example.org/api/contact",
  "user_info_fields": ["name", "email", "message"],
  "requires_consent": true
}

🗓 booking_slots

Expose appointment or availability information for scheduling.

"booking_slots": {
  "endpoint": "https://example.org/api/slots",
  "available_hours": ["2025-06-01T10:00Z", "2025-06-01T14:00Z"],
  "duration_minutes": 30,
  "fields_required": ["name", "email", "notes"]
}

🌍 regional_presence

Declare where and how your service is available geographically.

"regional_presence": {
  "region": "fr",
  "languages": ["fr", "en"],
  "support": {
    "available": true,
    "contact": "mailto:support@example.org"
  }
}

These blocks are optional but highly recommended for services with real-world contact points or international reach.

Relation to other feeds

Feed Purpose
llm-index.llmfeed.json Lists this and others for indexing
capabilities.llmfeed.json May be separated for scalability or auth
manifesto.llmfeed.json Declares the long-term intention or ethics
mcp-lite.llmfeed.json A compressed summary for low-bandwidth devices

Best Practices

  • Use minimal prompts[] and route them clearly
  • Separate advanced capabilities if access-scoped
  • Always include a trust block, even if unsigned
  • Reference from your llm-index.llmfeed.json

Related

🧠 intent_router (optional)

The intent_router block allows an MCP feed to declare what kinds of user intentions (natural language prompts) an agent should be ready to respond to, and what it should suggest in return.

Purpose

This turns the MCP feed into a router of human intent, giving the agent clear guidance on what to say or propose if the user asks common questions like "What is this?", "How can I join?", or "Is this feed certified?".

Example

"intent_router": [
  {
    "intent": "educate_user_about_mcp",
    "triggers": [
      "what is mcp",
      "how does this work",
      "explain model context protocol",
      "teach me"
    ],
    "response_hint": "Explain briefly what MCP is and offer to open either the spec or the compiled-site feed.",
    "suggested_targets": [
      {
        "title": "compiled-site.llmfeed.json",
        "url": "https://wellknownmcp.org/.well-known/exports/compiled-site.llmfeed.json"
      },
      {
        "title": "spec.llmfeed.json",
        "url": "https://wellknownmcp.org/.well-known/spec.llmfeed.json"
      }
    ]
  }
]

Fields

Field Required Description
intent ✅ Canonical name of the detected user intent
triggers[] ✅ Phrases (or regex) that the LLM should match
response_hint ✅ Short description of what the agent should say/do
suggested_targets[] ✅ Pages or feeds the LLM can suggest or load

This field complements the prompts[] block, but works in reverse: rather than giving a prompt to the LLM, it tells the LLM how to reply to a user prompt.