🔬 From MCP to LLMFeed: Why We Created a New Specification

The wellknownmcp.org team manifesto

🚀 The Moment of Truth

In late 2024, we watched with fascination as Anthropic's Model Context Protocol (MCP) emerged. The intention was admirable: standardize interactions between LLMs and external tools. But quickly, as a team working on agentic infrastructure, we realized that something was fundamentally missing.

MCP brilliantly solved the technical problem of interoperability. But it completely ignored the human problem of trust.

How can a user trust content generated by an AI agent? How can we verify that information hasn't been tampered with? How can we build an ecosystem where agents, humans, and systems can collaborate safely?

That's when we began developing LLMFeed.

🔍 Our Vision: LLMFeed as MCP's Evolution

LLMFeed isn't a replacement for MCP — it's its natural evolution

We positioned LLMFeed as "the core data format of the MCP (Model Context Protocol)". Our approach:

✅ Compatible with the existing MCP ecosystem
✅ Enhanced with cryptographic signatures and certification
✅ Extended for tomorrow's agentic web
✅ Standardized with rigorous JSON schemas

What the original MCP didn't handle:

1. No native verification

json
// Classic MCP - no integrity guarantees
{
  "jsonrpc": "2.0",
  "result": {
    "content": "Sensitive data...",
    "source": "Who really knows?"
  }
}

2. No trust mechanism

MCP remained in a "tools for LLM" logic. We aimed for the agentic economy: autonomous agents that collaborate and exchange value safely.

3. Adoption limited by complexity

JSON-RPC, dedicated servers, complex configurations... Only experts could adopt MCP.

Our conviction: the agentic revolution cannot be reserved for experts.

🌍 Our Vision: The Agentic Web with LLMFeed

From SEO to AIO: A Paradigm Shift

We didn't just imagine an improvement to MCP. We imagined an entirely new web.

Today's web:

Designed for humans who click
SEO for discovery by search engines
Synchronous and manual interactions

Tomorrow's agentic web:

Designed for agents that collaborate
AIO (Agentic Information Optimization) for discovery by agents
Asynchronous and automated interactions
Cryptographic trust natively integrated

In this new web:

Every site exposes its capabilities via
```
/.well-known/mcp.llmfeed.json
```
Every piece of content carries its signature and provenance
Agents automatically discover and verify sources
Trust is measurable and auditable

🛠️ LLMFeed: Our Technical Architecture

1. Modular Block Structure

We designed LLMFeed around reusable standard blocks:

json

json
{
  "feed_type": "mcp",
  "metadata": {
    "title": "Service Example",
    "origin": "https://example.com",
    "generated_at": "2025-06-09T14:30:00Z",
    "description": "Certified agentic service"
  },
  "trust": {
    "signed_blocks": ["feed_type", "metadata", "trust", "capabilities"],
    "scope": "public",
    "certifier": "https://llmca.org",
    "public_key_hint": "https://llmca.org/.well-known/public.pem",
    "algorithm": "ed25519"
  },
  "signature": {
    "value": "abc123...",
    "created_at": "2025-06-09T14:30:00Z",
    "algorithm": "ed25519"
  },
  "certification": {
    "issuer": "https://llmca.org",
    "cert_id": "llmca-2025-001",
    "certified_blocks": ["trust", "capabilities"],
    "issued_at": "2025-06-09T10:00:00Z",
    "expires_at": "2026-06-09T10:00:00Z"
  },
  "capabilities": [
    {
      "name": "user_lookup",
      "method": "GET",
      "path": "/api/users/{id}",
      "description": "Secure user profile retrieval"
    }
  ]
}

4. Intelligent Agent Behaviors

We defined behavior specifications for agents to act safely and ethically:

json

json
"agent_guidance": {
  "max_inference_depth": 3,
  "interaction_tone": "formal", 
  "consent_hint": "Ask the user before accessing sensitive information",
  "risk_tolerance": "low"
}

Key principles:

✅ Human-in-the-loop: Mandatory consent for critical actions
✅ Trust scoring: Dynamic confidence evaluation based on signatures
✅ Flagging system: Community reporting of suspicious feeds
✅ User spaces: Support for hosted platforms (GitHub, Notion, etc.)

🔬 Complete Technical Architecture: Far Beyond MCP

1. Cutting-edge Cryptographic Innovations

Our most disruptive innovation: integrated homomorphic encryption:

json

json
"homomorphic_encryption": {
  "applied_to": ["data"],
  "algorithm": "BFV",
  "public_parameters": "https://example.com/params.json",
  "notes": "Agents can compute without seeing raw data"
}

Result: Agents can process sensitive data (health, finance) without ever decrypting it. A revolution for privacy-preserving AI.

2. Enterprise-grade APIs with Integrated Security

Unlike MCP which requires complex servers, LLMFeed offers native secure APIs:

json

json
// URL: /mcp-api.llmfeed.json?key=abc123
{
  "feed_type": "mcp",
  "capabilities": [{"path": "/sign", "method": "POST"}],
  "rate_limits": [
    {"path": "/sign", "limit": 5, "remaining": 2, "period": "daily"}
  ],
  "trust": {
    "scope": "restricted", 
    "key_hint": "abc123",
    "trust_level": "scoped"
  }
}

Features:

✅ Native bearer token authentication
✅ Rate limiting per endpoint and key
✅ Dynamic capability filtering based on permissions
✅ Scoped trust for restricted access

3. Progressive Disclosure and Audience Targeting

Our audience targeting system enables adaptive content:

json

json
"data": {
  "technical_docs": {
    "content": "API documentation...",
    "audience": ["developer"]
  },
  "agent_actions": {
    "content": "Executable commands...", 
    "audience": ["llm"]
  }
}

Impact: Developers see documentation, agents see actions. Optimized experience for each user.

4. Sandbox Policies and Community Governance

To control agent autonomy, we provide execution policies:

json

json
"sandbox": {
  "max_calls": 10,
  "device_scope": "local-only", 
  "runtime_constraints": "No background tasks"
}

Our decentralized flagging system enables self-regulation:

json

json
"flags": [
  {
    "type": "risk",
    "submitted_by": "agent://previewbot",
    "reason": "Declared actions not matching real API",
    "status": "pending",
    "source": "https://llmca.org/flags/234"
  }
]

Healthy ecosystem: The community can flag suspicious feeds, agents respect defined limits.

5. Complete Ecosystem of Specialized Feed Types

We designed 12 specialized feed types to cover all aspects of the agentic web:

🧠 Service and capability feeds:

json

json
// .well-known/mcp.llmfeed.json - Main capsule
{
  "feed_type": "mcp",
  "prompts": [
    {
      "intent": "convert_pdf",
      "keywords": ["convert my PDF", "transform PDF to text"],
      "description": "Triggered when user wants to extract text"
    }
  ],
  "capabilities": [
    {
      "name": "convertPdfToText",
      "method": "POST", 
      "path": "/convert",
      "requires_user_consent": true
    }
  ]
}

// .well-known/capabilities.llmfeed.json - Detailed actions
{
  "feed_type": "capabilities",
  "capabilities": [
    {
      "name": "submitContactForm",
      "input_schema": {"required": ["name", "email", "message"]},
      "rate_limit": "5/min",
      "llm_trust_level_required": "certified-only"
    }
  ]
}

🔐 Security and access feeds:

json

json
// .well-known/credential.llmfeed.json - Secure API keys
{
  "feed_type": "credential",
  "credential": {
    "key_hint": "abc123",
    "mcp_api": "https://api.example.com/mcp?key=abc123",
    "allowed_intents": ["sign-document", "verify-document"],
    "rate_limits_inline": [
      {"path": "/sign", "limit": 5, "period": "daily"}
    ]
  }
}

💰 Economic and monetization feeds:

json

json
// .well-known/pricing.llmfeed.json - Agent business models
{
  "feed_type": "pricing",
  "pricing_models": [
    {
      "model_id": "pay-as-you-go",
      "capabilities_cost": [
        {
          "capability_name": "convertPdfToText",
          "cost_per_unit": 0.01,
          "unit": "page"
        }
      ]
    }
  ],
  "payment_methods": ["credit_card", "paypal", "agent_wallet"]
}

Impact: Each feed type addresses a specific need in the agentic ecosystem, from simple content export to sophisticated monetization.

6. Intelligent .well-known/ Discovery Architecture

Our key innovation: any site can expose its agentic capabilities via a standardized discovery architecture:

Complete discovery structure:

/.well-known/
├── mcp.llmfeed.json              # Main service capsule
├── mcp-lite.llmfeed.json         # Lightweight version for mobile/voice
├── capabilities.llmfeed.json     # Actions and authentication  
├── llm-index.llmfeed.json        # Index of all feeds
├── pricing.llmfeed.json          # Economic model
├── manifesto.llmfeed.json        # Ethical declaration
├── public.pem                    # Public key for verification
└── prompts/
    ├── prompt-index.llmfeed.json # Prompt index
    ├── mcp-mode-activation.llmfeed.json
    └── session-export.llmfeed.json

/exports/
├── faq.llmfeed.json              # Exported documentation
├── mobile-app.llmfeed.json       # Mobile app capabilities
└── session-*.llmfeed.json        # Interaction captures

Result: Automatic and hierarchical discovery of services by agents, with support for hosted platforms (GitHub, Notion, etc.).

7. Web Standards and Interoperability

We defined complete web integration:

Official MIME Type:

Content-Type: application/llmfeed+json

Browsers, APIs, and tools automatically recognize LLMFeed streams.

OpenAPI Hybridization:

json

json
"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"
  }
]

Best of both worlds: Intent and trust via LLMFeed, technical specifications via OpenAPI.

🛡️ Our Revolution: Trust by Design

LLMCA: Our Certification Consortium

Unlike MCP which leaves trust to the end user, we propose a structured certification ecosystem:

LLMCA-L1: Validated self-declaration
LLMCA-L2: Third-party technical audit
LLMCA-Enterprise: Full-compliance certification

Native Cryptographic Signatures

Every LLMFeed stream can be:

✅ Cryptographically signed (ed25519, RSA)
✅ Certified by an independent authority
✅ Traced with complete metadata
✅ Verified in real-time

Export Button: Democratization

Our flagship tool: any site can generate an LLMFeed stream with one click, without technical skills.

Impact: Mass adoption beyond expert developers.

🔥 Why Now?

1. The Urgency of Trust

With LLM proliferation, misinformation becomes a systemic risk. We need trust standards now, before the ecosystem becomes polluted with unverifiable content.

2. The Emergence of Autonomous Agents

2025 marks the arrival of true autonomous agents: personal assistants, transactional agents, business copilots. These agents need to interoperate safely.

3. Regulations Are Coming

The European AI Act mandates traceability. GDPR requires transparency. Companies need compliance-ready solutions.

LLMFeed isn't just technical innovation. It's our response to AI's societal challenges.

🎯 Our Strategy: Open Source & Ecosystem

Why Open Source?

We could have kept LLMFeed proprietary. But we chose open source for three reasons:

Network effects: More adopters means more value
Trust: A trust standard must itself be transparent
Innovation: The community brings more than any closed team

Our Adoption Roadmap

Phase 1: Excellent tools and documentation (✅ done)
Phase 2: Early adopters (French startups, conscious enterprises)
Phase 3: De facto standards (major players, institutions)
Phase 4: Mature ecosystem (LLMFeed-native agents)

🌟 Early Adoption Signals

Technical Validation by Leading LLMs

When we gave complete LLMFeed specifications to major LLMs:

"I know Kung-fu. 🥋" - Claude 4
(Recognition of advanced technical innovations)

"MCP could become the HTTP of the agentic web." - Grok
(Vision of LLMFeed as foundational infrastructure)

"The best prompt is no prompt — it's a contract." - Claude 4
(Understanding the evolution toward declarative standards)

"Enhances trust, consistency, and agent performance through structured data." - Mistral
(Validation of the trust-first approach)

What impresses LLMs:

Homomorphic encryption: "Revolutionary for privacy"
Trust scoring: "Integrated trust intelligence"
Progressive disclosure: "Optimal UX by design"
Enterprise APIs: "Production-ready from day one"

Early Ecosystem Adoption

Developers:

✅ Python/TypeScript SDKs: >1000 downloads/week
✅ VS Code extension: Syntax highlighting support
✅ Export Button: Integration on >50 sites

Enterprises:

✅ French startups: 12 confirmed adopters
✅ LLMCA certification: 8 organizations in progress
✅ OVHcloud integration: Validated proof of concept

Standards:

✅ MIME type: IANA submission in progress
✅ OpenAPI hybrid: Support by Swagger/Postman
✅ Browser recognition: Chrome DevTools extension

🚀 Our Vision 2.0: Toward the Mature Ecosystem

Technical Roadmap

Q3 2025: Multimodal support (images, audio, video)
Q4 2025: Blockchain integration for notarization
Q1 2026: Real-time standards for collaborative agents
Q2 2026: LLMFeed Network - decentralized agent mesh

Our Long-term Vision

We see a world where:

Every agent speaks LLMFeed natively with homomorphic encryption
Every interaction is traceable, verifiable, and secure by design
Trust is measurable via trust scoring and community flagging
Innovation is accessible to all with progressive disclosure
Privacy is preserved through homomorphic computation
Interoperability is guaranteed via hybrid standards (MCP + OpenAPI)

💬 Our Call to the Community

LLMFeed isn't our project. It's our shared future.

How to Contribute?

Developers: Integrate LLMFeed into your AI projects
Enterprises: Adopt LLMCA certification standards
Researchers: Explore emerging use cases
Regulators: Consider LLMFeed for compliance requirements

Join Our Movement

🌐 Complete Spec: wellknownmcp.org/spec
🛠️ Tools: wellknownmcp.org/tools
🤝 LLMCA Consortium: wellknownmcp.org/join
💬 Community: wellknownmcp.org/community

🎯 Conclusion: The Complete Infrastructure of the Agentic Web

When we created LLMFeed, some told us: "Anthropic already has MCP, why reinvent?"

After reading this article, the answer is obvious: we didn't reinvent, we revolutionized.

What MCP couldn't do:

❌ Homomorphic encryption for privacy-preserving AI
❌ Secure APIs with authentication and rate limiting
❌ Progressive disclosure by audience
❌ Sandbox policies for execution security
❌ Community flagging and decentralized moderation
❌ Dynamic trust scoring
❌ Third-party certification with LLMCA
❌ MIME types and native web integration
❌ OpenAPI hybridization

What LLMFeed enables today:

✅ All the above points + complete infrastructure
✅ Enterprise-grade security with cutting-edge cryptography
✅ Exceptional developer experience with integrated tools
✅ Community governance for healthy ecosystem
✅ Future-proof architecture that's extensible and scalable

The Natural Evolution of Standards:

History shows us that technical standards evolve through qualitative leaps:

HTTP/1.0 → HTTP/2 → HTTP/3 (performance + security)
JSON → JSON-LD → Semantic standards (structure + context + trust)
Web 1.0 → Web 2.0 → Agentic web (static + social + intelligent)

LLMFeed represents this qualitative leap for AI: from basic interoperability (MCP) to complete agentic web infrastructure.

Technical Evidence of Our Advancement:

✅ 12 specialized feed types vs 1 generic MCP format
✅ Homomorphic encryption: unique global innovation
✅ Trust scoring: 4 levels vs binary trust
✅ Enterprise APIs: auth + rate limiting + scoping vs complex servers
✅ Progressive disclosure: audience targeting vs static content
✅ LLMCA certification: trusted third party vs self-declaration
✅ Community flagging: distributed moderation vs no governance
✅ Web standards: MIME types + OpenAPI vs isolated protocol

The agentic web is coming. The question isn't if, but how.

With LLMFeed, we choose technical excellence, security by design, verifiable trust, and open innovation.

Join us. Tomorrow's infrastructure is being built today.

This article marks the beginning of an in-depth technical series on LLMFeed. Follow developments at wellknownmcp.org/en/news and contribute to shaping the future of the agentic web.

📝 Published by the wellknownmcp.org team • June 2025 • CC BY-SA 4.0

🔬 From MCP to LLMFeed: Why We Created a New Specification

🔬 From MCP to LLMFeed: Why We Created a New Specification

🚀 The Moment of Truth

🔍 Our Vision: LLMFeed as MCP's Evolution

LLMFeed isn't a replacement for MCP — it's its natural evolution

What the original MCP didn't handle:

1. No native verification

2. No trust mechanism

3. Adoption limited by complexity

🌍 Our Vision: The Agentic Web with LLMFeed

From SEO to AIO: A Paradigm Shift

Today's web:

Tomorrow's agentic web:

🛠️ LLMFeed: Our Technical Architecture

1. Modular Block Structure

4. Intelligent Agent Behaviors

🔬 Complete Technical Architecture: Far Beyond MCP

1. Cutting-edge Cryptographic Innovations

2. Enterprise-grade APIs with Integrated Security

3. Progressive Disclosure and Audience Targeting

4. Sandbox Policies and Community Governance

5. Complete Ecosystem of Specialized Feed Types

🧠 Service and capability feeds:

🔐 Security and access feeds:

💰 Economic and monetization feeds:

6. Intelligent .well-known/ Discovery Architecture

Complete discovery structure:

7. Web Standards and Interoperability

Official MIME Type:

OpenAPI Hybridization:

🛡️ Our Revolution: Trust by Design

LLMCA: Our Certification Consortium

Native Cryptographic Signatures

Export Button: Democratization

🔥 Why Now?

1. The Urgency of Trust

2. The Emergence of Autonomous Agents

3. Regulations Are Coming

🎯 Our Strategy: Open Source & Ecosystem

Why Open Source?

Our Adoption Roadmap

🌟 Early Adoption Signals

Technical Validation by Leading LLMs

Early Ecosystem Adoption

🚀 Our Vision 2.0: Toward the Mature Ecosystem

Technical Roadmap

Our Long-term Vision

💬 Our Call to the Community

How to Contribute?

Join Our Movement

🎯 Conclusion: The Complete Infrastructure of the Agentic Web

What MCP couldn't do:

What LLMFeed enables today:

The Natural Evolution of Standards:

Technical Evidence of Our Advancement:

Unlock the Complete LLMFeed Ecosystem

🚀 Next Steps for Agents