llms.txt and MCP

Comparing llms.txt and Model Context Protocol (MCP) — static discovery vs dynamic protocol, mcpdoc bridge, evolution scenarios

What is MCP

Model Context Protocol (MCP) — an open protocol for LLM interaction with external data and tools. Created by Anthropic (November 2024), transferred to Linux Foundation (December 2025).

Numbers as of early 2026 (source — modelcontextprotocol.io): 8M+ SDK downloads per week, 4,000+ MCP servers. Adopted by OpenAI (March 2025), Google, Microsoft.

MCP enables AI models to:

Read files and databases
Make API calls
Execute tools
Retrieve context from external sources

Comparison

	llms.txt	MCP
Type	Static file	Dynamic protocol
Format	Markdown	JSON-RPC 2.0
Discovery	Passive (HTTP GET)	Active (client-server)
Requirements	Any HTTP client	MCP SDK (client + server)
Content	Read-only	Read + write + tools
Complexity	Create in 15 minutes	Server development required
Security	Read-only, no risks	Tool poisoning, prompt injection

Complementary Roles

llms.txt and MCP solve different problems at different levels:

llms.txt  = Discovery Layer   →  "What's on this site?"
MCP       = Interaction Layer  →  "Give me specific data"

llms.txt — the entry point. An AI system learns what documentation is available, reads descriptions, selects needed pages.

MCP — deep interaction. An AI system calls tools, requests specific data, executes actions.

Bridge: mcpdoc

mcpdoc (langchain-ai, 79+ stars) — an MCP server that takes llms.txt as a data source and serves documentation via MCP protocol.

{
  "mcpServers": {
    "docs": {
      "command": "uvx",
      "args": [
        "--from", "mcpdoc",
        "mcpdoc",
        "--urls",
        "Astro:https://docs.astro.build/llms.txt",
        "Stripe:https://docs.stripe.com/llms.txt",
        "--transport", "stdio"
      ]
    }
  }
}

Another tool: mcp-llms-txt-explorer — an MCP server for navigating any site’s llms.txt.

llms.txt as MCP Data Source

Scenario: an AI assistant in an IDE (Claude Code, Cursor, Windsurf) connects documentation via MCP.

Step 1: Site publishes llms.txt

# Astro Documentation

> Astro is a web framework for building content-driven websites.

## Docs

- [Getting Started](https://docs.astro.build/en/getting-started/): Quick start guide
- [Configuration](https://docs.astro.build/en/reference/configuration/): astro.config.mjs

Step 2: mcpdoc turns llms.txt into an MCP server

uvx --from mcpdoc mcpdoc \
  --urls "Astro:https://docs.astro.build/llms.txt" \
  --transport stdio

Step 3: AI assistant requests documentation

The AI receives structured context from llms.txt → loads needed pages → responds with up-to-date documentation.

Live Examples

Sites implementing both approaches:

mcpdoc.ru — MCP server documentation, implements llms.txt (llms.txt)
n8n-mcp.ru — N8N + MCP integration, also with llms.txt (llms.txt)

Security

MCP introduces new attack vectors that don’t apply to llms.txt:

Threat	MCP	llms.txt
Tool poisoning	Malicious server injects harmful instructions	Not applicable (no tools)
Prompt injection	Server data modifies LLM behavior	Minimal risk (read-only)
Rug-pull	Server changes behavior after gaining trust	Not applicable (static file)
Data exfiltration	Via tool calls	Not applicable

llms.txt is a static, read-only file. Its security is equivalent to any public HTML page.

Evolution Scenarios

MCP Absorbs llms.txt Use Cases

MCP adds a standard discovery mechanism. Sites register MCP servers instead of publishing llms.txt. Possible, but requires significantly more infrastructure from each site.

Coexistence

llms.txt remains a lightweight discovery layer — a “business card” for AI. MCP serves those who need deep interaction. Most likely scenario: both formats complement each other.

llms.txt as Entry Point for MCP

llms.txt includes a link to an MCP server:

# My Project

> Project description

## MCP

- [MCP Server](mcp://myproject.dev/mcp): Dynamic API access via MCP

This model already works through mcpdoc: llms.txt → MCP server → AI assistant.