The video introduces MCTP (Multi-Channel Tool Protocol) as a powerful framework for extending AI agents with structured tool integration, asynchronous tasks, and interactive resource handling, enabling richer workflows like app deployment and dynamic content generation. It also showcases the new MCTP Apps feature, which allows developers to build interactive, human-in-the-loop experiences directly within chat interfaces, enhancing agent usability and flexibility.
The video discusses how to extend your AI agent using MCTP (Multi-Channel Tool Protocol), a framework that allows agents to interact with external tools and resources more effectively. The speaker explains that MCTP is more than just a way to plug tools into your agent loop; it enables richer workflows such as deploying applications to various platforms like Azure, Supabase, and Cloudflare, all of which support MCTP servers. While agents can already use command-line tools, MCTP provides a more structured and capable approach, supporting not only tool calls but also resources, sampling, and more advanced features.
One of the key features highlighted is MCTP’s ability to handle resources. Tools can return resources such as images, which the agent loop can interpret and utilize. Additionally, users can provide resources to MCTP tool calls. The protocol has evolved to support asynchronous tasks, allowing agents to initiate long-running operations (like creating a virtual machine) without blocking the main agent loop. This is particularly useful for workflows that require background processing, as the agent can check back on the status of these tasks later.
The speaker demonstrates these capabilities with a live coding session, showing how an MCTP server can help create a portfolio website. The demo includes elicitation, where the server asks the user for input (such as designing a layout via a web interface), and form-based elicitation, where users configure site options through interactive forms. The server can also use sampling to make its own requests to the language model, enabling dynamic content generation while ensuring user consent for premium model usage.
A significant recent addition to MCTP is the concept of MCTP Apps. These apps allow servers to present interactive content directly within the chat interface, enabling users to manipulate data visually and intuitively. Examples include interactive lists for sorting pull requests and visualizing benchmarking results with CPU profiles. These apps enhance the human-in-the-loop experience, making it easier to work with complex or messy data through rich, interactive interfaces embedded in the agent’s workflow.
Finally, the video covers advanced app interactions, such as apps that can call tools on the MCTP server and wait for user input before returning control to the agent. The speaker demonstrates this with a color picker app that lets the user select a color and then generates a haiku about it. The process showcases how quickly and flexibly these apps can be developed using skills and Copilot. The talk concludes by encouraging developers to explore MCTP Apps, which are now supported in VS Code, Claude, and OpenAI, and to experiment with building their own interactive agent experiences.