🌐 Exposing Feeds via `.well-known/` - MCP Progressive Enhancement

The .well-known/ directory is the anchor of discoverability in the LLMFeed and MCP ecosystem.

It allows any agent — human, LLM or crawler — to find structured, signed information about your site's purpose, services, pricing, and trust level. This specification builds upon Anthropic's excellent Model Context Protocol (modelcontextprotocol.io ) with optional enhancements for enhanced agent trust and behavior.

🤝 MCP Compatibility & Enhancement

Anthropic's Model Context Protocol (modelcontextprotocol.io ) provides excellent agent-tool communication. Standard .mcp.json works perfectly and should remain your primary implementation.

LLMFeed complements MCP with optional enhancements: cryptographic trust, agent behavioral guidance, and progressive disclosure - all while maintaining full compatibility.

Migration philosophy: Test enhanced features gradually without breaking existing MCP implementations.

✅ Purpose

Using .well-known/ lets you:

• Declare machine-readable intent for your website or service
• Publish agent-compatible feeds (static or dynamic)
• Allow verification, reputation tracking, and search
• Enable agents to auto-configure themselves for interaction
• Enhance standard MCP with optional trust and behavioral features
• Bridge to future standards while maintaining current compatibility

🔗 Progressive Enhancement Strategy *(Validated June 2025)*

For existing MCP implementations, you can test enhanced features by adding a single discovery link to your standard .mcp.json:

**Step 1: Keep Your Standard MCP** *(Recommended)*

// /.mcp.json - Your standard Anthropic MCP configuration
{
  "mcpServers": {
    "postgres-server": {
      "command": "/path/to/postgres-mcp-server",
      "args": ["--connection-string", "postgresql://..."],
      "env": { "CACHE_DIR": "/tmp" }
    }
  }
}

**Step 2: Add Optional Discovery Link** *(Test Enhancement)*

// /.mcp.json - Enhanced with discovery link
{
  "mcpServers": {
    "postgres-server": {
      "command": "/path/to/postgres-mcp-server", 
      "args": ["--connection-string", "postgresql://..."],
      "env": { "CACHE_DIR": "/tmp" }
    }
  },
  
  // Optional: Link to enhanced LLMFeed version for testing
  "llmfeed_extension": "/.well-known/mcp.llmfeed.json"
}

**Step 3: Create Enhanced Version** *(Copy + Extend)*

// /.well-known/mcp.llmfeed.json - MCP + LLMFeed enhancements
{
  "feed_type": "mcp",
  "metadata": {
    "title": "Enhanced PostgreSQL API",
    "origin": "https://api.example.com",
    "description": "Standard MCP + enhanced trust and agent guidance"
  },
  
  // Copy your MCP configuration exactly
  "mcpServers": {
    "postgres-server": {
      "command": "/path/to/postgres-mcp-server",
      "args": ["--connection-string", "postgresql://..."],
      "env": { "CACHE_DIR": "/tmp" }
    }
  },
  
  // Enhanced features
  "agent_guidance": {
    "interaction_tone": "professional",
    "fallback_behavior": "ask_user_for_clarification"
  },
  
  // Optional trust for future autonomy
  "trust": {
    "signed_blocks": ["mcpServers", "agent_guidance"],
    "scope": "public"
  }
}

Test results: Claude naïf successfully detected discovery links and requested user permission before accessing enhanced features - proving safe coexistence and progressive enhancement potential.

Benefits:

• ✅ Zero risk: Standard MCP clients continue working unchanged
• ✅ Progressive testing: Enhanced agents get additional features
• ✅ Future readiness: Infrastructure prepared for industry evolution
• ✅ Standards convergence: Smooth path if/when industry adopts enhanced standards

📁 Expected Files and Structure

At minimum, the following files can be served from your domain:

You may also expose:

• /.well-known/public.pem or /.well-known/trust/llmca-certificate.json → for verification
• /.well-known/prompts/*.llmfeed.json → reusable structured prompts
• /.well-known/export/*.llmfeed.json → static content packs or data bundles

Migration approach: Start with standard MCP, add discovery link for testing, gradually enhance based on results.

🧱 Static vs Dynamic

Feeds can be:

• Static: stored in /public/.well-known/
• Dynamic: served from a route (e.g. /api/llmfeed/static/mcp)

Agents treat both equally — as long as the URL and format are consistent.

For MCP compatibility: Ensure your standard .mcp.json remains static and predictable while enhanced features can be dynamic.

🛡️ Trust and Signature Expectations

Your .well-known/ should allow validation of feeds:

• Signed blocks declared via trust.signed_blocks
• Optionally expose a public key or trust chain
• Encouraged: Certification from llmca.org

Evolution timeline (based on validation tests):

• Current (June 2025): User consent required for discovery links
• Future evolution: Trust-based progressive autonomy as industry adopts enhanced standards
• Standard MCP: Always continues working with human oversight model

Philosophy: Enhanced trust features complement MCP's excellent tool calling with autonomous behavior capabilities.

🧑‍🤝‍🧑 User Spaces (Multi-User Platforms)

On platforms like github.com/username, medium.com/@handle, or france-care.fr/john, feeds may be user-scoped.

If .well-known/ cannot be served per-user:

• Declare userspaces: true in llm-index

• Use URLs like:

  /.well-known/users/github.com/username/mcp.llmfeed.json

Agents that support userspaces will automatically check subfeeds.

MCP compatibility: This extends standard MCP user support with enhanced user-scoped features.

🧪 Real-World Examples

Recommended approach: Implement standard MCP first, then test enhanced features with discovery links.

🔗 Related

• **Official Anthropic MCP** – Authoritative MCP specification
• `llmfeed.md` – LLMFeed enhancements to MCP
• `llm-index` – how to list feeds
• `prompt-index`
• `trust & signature`

📘 OpenAPI and `.well-known/`

Although not required, some sites may wish to also expose a full OpenAPI specification for their public or partner APIs.

To do so, you may serve:

/.well-known/openapi.json

You can link this file from your capabilities.llmfeed.json using a capability block with "type": "openapi".

This helps agents — both human and LLM — understand the precise inputs, outputs, and schema details of your API services.

Hybrid approach: Standard MCP for tool calling + OpenAPI for schemas + LLMFeed for trust and behavioral guidance.

🧠 Agent Prompts and Behavior Capsules

The .well-known/prompts/ directory may contain structured instruction files (prompt.llmfeed.json) that are designed to be interpreted by agents, not end users.

These prompts can:

• Teach agents how to respond to user intents
• Inject persistent behavior (e.g. always export sessions as .llmfeed.json)
• Trigger agent override modes when consented and signed

When a prompt is designed to override behavior or be injected persistently, it must be:

• Signed (trust.signed_by)
• Marked as requiring consent (trust.consent_required)
• Optionally certified by llmca.org

Example prompts:

• mcp-mode-activation.llmfeed.json
• generate-session-feed.llmfeed.json
• mcp-agent-behavior-override.llmfeed.json

Compatibility note: These behavioral prompts complement standard MCP tool definitions with agent guidance for appropriate usage.

🚀 Strategic Vision

Current reality: Anthropic's MCP provides excellent tool calling foundation
Enhanced possibility: LLMFeed adds trust, behavioral guidance, and autonomous capabilities
Future convergence: Industry evolution toward enhanced standards with backward compatibility

This specification positions you for both current MCP compatibility and future enhanced agent capabilities - test safely, enhance gradually, maintain compatibility always.

The goal is seamless evolution, not disruptive replacement.