Is MCP the Future of N8N AI Agents? (Fully Tested!)

Key Concepts

• mCP (Model Context Protocol): A standardized way for AI agents to discover and use tools and data sources.
• AI Agent Host: The application hosting the AI agent (e.g., n8n, Claude desktop).
• mCP Client: A component within the AI agent host that connects to mCP servers.
• mCP Server: A service that exposes tools and data sources in a standardized way.
• Tools: Actions or functionalities offered by a service (e.g., scraping a webpage, searching the web).
• Resources: Data or content exposed by a service (e.g., file contents, database records).
• Prompts/Prompt Templates: Instructions or examples provided by the mCP server to guide the AI agent in using the tools effectively.
• Transports (STDIO/SSE): Methods for communication between the mCP client and server (Standard Input/Output, Server-Sent Events).
• Sampling: A mechanism for human-in-the-loop interaction, allowing the server to request user confirmation or rejection of actions.
• Routes: Boundaries or scopes defined within the service (e.g., projects, repositories).

mCP Overview and Motivation

• mCP is an open standard published by Anthropic, initially motivated by the need for their Claude desktop application to interact with local and remote services.
• It aims to provide a "USB for AI models," enabling seamless integration of software applications into AI agents.
• The core idea is to abstract away the specifics of each tool and service, allowing agents to discover and use them dynamically.
• The video highlights an n8n community module that allows users to experiment with mCP.

Traditional n8n Agent Tools vs. mCP

• Traditional Approach: Requires manually creating specific tools for each action and hardcoding instructions in the system prompt. This is consuming and doesn't scale well.
• mCP Approach: Uses a single endpoint to list available tools from an mCP server. The agent can then execute those tools.
• Key Advantage of mCP: As tools are added or improved on the server side, the agent automatically gains access to them without requiring modifications to the n8n workflow.
• Trade-offs: mCP adds complexity due to the additional layer of the mCP server. Security concerns arise regarding authentication and authorization.

mCP Architecture and Workflow

1. AI Agent Host (n8n): Contains the AI agent and the mCP client.
2. mCP Client: Connects to the mCP server.
3. mCP Server: Acts as a bridge between the client and the underlying service (e.g., Fir Crawl, Brave Search).
4. Service API: The API of the service that the mCP server interacts with.
5. Communication Flow:
   • Agent requests a list of available tools from the mCP server.
   • Server responds with the list of tools and their descriptions.
   • Agent selects a tool and provides the necessary parameters.
   • Server executes the tool and returns the result to the agent.

Setting up mCP Clients and Servers in n8n

• The video demonstrates how to set up mCP clients and servers using the n8n community module.
• Installation: The community module needs to be installed within the n8n instance (self-hosted, not available in n8n Cloud).
• Server Configuration: The mCP server can be run either permanently or temporarily using npx. The video uses npx to spin up temporary servers for Fir Crawl and Brave Search.
• Client Configuration: The mCP client needs to be configured with the command to start the server and any necessary API keys.
• Transport Options: STDIO (Standard Input/Output) for local communication and SSE (Server-Sent Events) for remote communication.

Examples and Demonstrations

• Fir Crawl: The video demonstrates scraping a webpage using the Fir Crawl mCP server.
• Brave Search: The video attempts to use the Brave Search mCP server to search for restaurants but encounters API limits.
• GitHub: The video shows the GitHub mCP server, which provides a list of resources (browser console logs) and tools (17 different tools).
• Puppeteer: The video attempts to use a Puppeteer agent for browser automation but fails due to permission issues within a Docker container.
• Apify: The video describes an attempt to connect to the Apify mCP server using SSE, but it was unsuccessful.

Key Features and Concepts Explained

• Tools: The mCP server exposes a list of available tools, along with their descriptions and required parameters.
• Prompts: The mCP server can provide prompt templates to guide the AI agent in using the tools effectively. This is considered a significant potential benefit of mCP.
• Resources: The mCP server can expose data and content from the service, such as file contents or database records.
• Transports: STDIO is suitable for local communication, while SSE is used for remote communication with the mCP server.
• Sampling: Allows for human-in-the-loop interaction, where the server can request user confirmation or rejection of actions.
• Routes: Define the boundaries or scopes of the service.

Comparison to n8n Agent Tools

• Overlap: Both agent tools and mCP can trigger actions within a service.
• mCP: Suited for general-purpose agents that need to access a wide range of tools.
• Agent Tools: Provide more control over the agent, allowing you to pick and choose specific tools and define their permissions.
• mCP Challenges: Backwards compatibility and repeatability of actions can be concerns, as new tools added to the server may change the agent's behavior.

Hal 90001 Example

• The video references a previous project, Hal 90001, a multi-agent n8n system with 26 sub-agents.
• mCP is seen as a potential solution for managing such complex systems, as it would allow the agents to dynamically discover and use tools from various services.

Challenges and Limitations

• Reliability: The video demonstrates that mCP can be unreliable, with issues such as incorrect parameter formatting and API limits.
• Documentation: The lack of detailed documentation for the tools can make it difficult for the AI agent to use them effectively.
• Maturity: The n8n community module and the mCP standard are still relatively new and under development.

Conclusion

• mCP has significant potential to simplify the integration of software applications into AI agents.
• However, it is not yet ready for widespread adoption due to issues with reliability, documentation, and maturity.
• The video encourages viewers to experiment with mCP and provide feedback to help improve the standard and the n8n community module.
• The presenter believes that mCP or a similar standard is needed to enable seamless integration of software applications into AI agents without requiring constant reinvention.