🔐 API Access for Agents

How agents discover what they can do using a credential and a feed

MCP makes the web agent-readable — without rebuilding it.

Large models like Claude or GPT increasingly support tool usage and external APIs. The Anthropic Model Context Protocol (MCP) proposes a way to expose those tools — but often assumes a “client SDK” architecture.

The LLMFeed approach extends that vision. Every website can now declare its own agent-compatible API surface, without requiring custom client code, user downloads, or central orchestration.

With a simple .llmfeed.json file, an agent can:

  • Understand the available actions
  • Verify permissions or limits
  • Automatically configure its intent-handling

The /mcp-api.llmfeed.json feed is how an agent retrieves its scoped access view once it receives a credential.

How it works

  • You give your agent a apicredential.llmfeed.json
  • That credential includes a mcp_api URL
  • The agent queries the endpoint with the key
  • Receives a signed feed with only the capabilities it can use
  • Reads rate_limits, prompts, trust scope
  • Acts or requests user confirmation

Example Feed

{
  "feed_type": "mcp",
  "capabilities": [
    { "path": "/sign", "method": "POST", "description": "Sign document" }
  ],
  "prompts": [
    { "intent": "sign-document", "description": "Prompt to trigger signing" }
  ],
  "rate_limits": [
    { "path": "/sign", "period": "daily", "limit": 5, "remaining": 2 }
  ],
  "trust": {
    "scope": "restricted",
    "key_hint": "abc123",
    "certifier": "https://llmca.org",
    "signed_blocks": ["capabilities", "prompts", "rate_limits", "trust"]
  }
}

Related pages