What is a LLMFeed?

📦 What is a .llmfeed.json?

A LLMFeed is a lightweight, signed and structured JSON file that lets any agent — including LLMs, copilots, and bots — understand what a website, document, or service is about and how to interact with it.

It is the core data format of the MCP (Model Context Protocol).

🧠 Why it matters

  • Machines need semantic context and trust — not just HTML or APIs
  • .llmfeed.json enables autonomous agents to:
    • Understand content
    • Detect intent
    • Act through capabilities
    • Evaluate cost or risk
    • Verify authorship and trust

🛠️ Minimum Structure

All valid feeds must include these two top-level blocks:

{
  "feed_type": "export",
  "metadata": {
    "origin": "https://example.org",
    "title": "Sample Page",
    "generated_at": "2025-05-20T15:00:00Z"
  }
}

Additional blocks are context-specific:

  • data, intent, trust, capabilities, prompts, session_state, credential, etc.

🧱 Standard Block Types

Block Purpose
feed_type Defines what kind of feed this is
metadata Describes title, origin, timestamp, etc.
trust Indicates signed blocks, trust scope
signature Optional hash and issuer info
certification Issued by trusted authority like LLMCA
data Carries exportable content (HTML, bundle, etc.)
intent Used in prompts, capabilities, or MCP
audience Optional target (e.g., llm, developer)

🧹 Common Feed Types

Feed Type Use Case Typical Blocks
mcp Describe your service/site metadata, intent, prompts, trust
export Share a page’s content with agents metadata, data, trust
prompt Structure prompts for reuse intent, audience, trust
session Reconstruct LLM session context metadata, session_state
credential Share API credentials or access claims metadata, scopes, rate_limits
pricing Describe monetization model pricing_models, payment_methods

Signatures and Certifications

To ensure the integrity and authenticity of a .llmfeed.json, it is RECOMMENDED to include a digital signature in the trust block, and list the signed_blocks that were covered by the signature.

Signing a feed is also a way to enforce the encapsulation of critical associations. It ensures that specific pieces of information — for example, an API key and its associated site, a quote and its author, or a payload and its intended context — are always processed together by consuming LLMs. This prevents unintentional detachment or misuse of key contextual elements.

Feeds may also be certified by trusted third parties (such as llmca.org) to provide an additional layer of verified trust. Certification helps ensure that consuming agents can rely on the feed's content and provenance.

Signed and certified feeds enable new models of interoperability, trust, and traceability across the Agentic Web.

🔐 Trust block

"trust": {
  "signed_blocks": ["feed_type", "metadata", "trust", "data"],
  "scope": "public",
  "certifier": "https://llmca.org",
  "public_key_hint": "https://llmca.org/.well-known/public.pem",
  "algorithm": "ed25519",
  "hints": "critical context integrity"
}

See `llmfeed_extensions_signatures.md`

🔏  signature block

"signature": {
  "value": "abc123...",
  "created_at": "2025-06-01T12:34:56Z"
}

🔁 Lifecycle

  1. Generated by user, script, platform or agent
  2. Optionally signed or certified
  3. Published (public URL or .well-known/)
  4. Parsed by agents for interaction
  5. May trigger actions, fallback, or ranking

📤 Export and Usage

It is RECOMMENDED to sign .llmfeed.json files to provide verification and trust for consuming LLMs.

Beyond serving feeds to crawlers or agents, .llmfeed.json files are particularly powerful as structured export formats. They enable what can be described as "copy-paste on steroids": any content, when wrapped in a signed and contextualized feed, can be shared, transferred, or reinjected into other agents or workflows while preserving its intended meaning and usage guidelines.

Such feeds can be used to:

  • Provide portable knowledge blocks.
  • Offer contextual copy/paste actions to LLMs or user interfaces.
  • Enable verified content sharing between sites or applications.
  • Maintain traceability and attribution through signatures.

The first usage of .llmfeed.json could be to create the Agentic Web, as feeds of any type (mcp, index, prompt, pricing, capabilities, api-credential, etc.) can be served by a website to provide value and context to any crawler, from simple robots to advanced LLMs.

Secured information exchange can drive the autonomy of agents and trigger self-discovered capabilities.

Feeds can be served from a site’s .well-known/ directory or included in APIs.

🗺️ Explore the full specification

🤝 MCP and OpenAPI

While capabilities[] provides a simplified way to describe callable functions and services, some developers may wish to include a link to a full OpenAPI specification.

This hybrid model allows agents to:

  • Understand the intent and trust level via MCP
  • Use OpenAPI for complete parameter definitions, schemas, and responses

Example:

"capabilities": [
  {
    "type": "endpoint",
    "intent": "get status",
    "url": "https://api.example.com/status"
  },
  {
    "type": "openapi",
    "url": "https://example.com/.well-known/openapi.json",
    "description": "Full technical spec for backend API"
  }
]

MCP and OpenAPI are complementary:

  • Use MCP for meaning, trust, and discovery.
  • Use OpenAPI for detailed technical implementation.