AIP-WG O. Daichi FOALK 15 January 2025 AI Transport Protocol draft-aipwg-ai-transport-latest Abstract AIP is designed to provide a framework for seamless integration of AI-powered communication systems, ensuring that AI agents can interact effectively with diverse web resources, facilitate efficient data exchanges, and ensure both factual accuracy and security within transportation protocols. This document outlines the key principles, concepts, and specifications for the AI Transport Protocol (AIP). Discussion Venues This note is to be removed before publishing as an RFC. Source for this draft and an issue tracker can be found at https://github.com/aip-wg/ai-transport. Table of Contents 1. Introduction 1.1. Motivation 1.1.1. Leveraging AI 1.1.2. Factual Integrity 1.1.3. Reachability 1.1.4. Accessibility 1.1.5. Traceability 1.1.6. Latency 1.2. Purpose 1.2.1. Large Integrated Resource 1.2.2. Educational Use of AI 1.3. Scope 1.4. Terms and Definitions 2. Concepts 2.1. Media over QUIC's Concepts 2.1.1. Transport 2.1.2. Session 2.1.3. Track 2.1.4. Group 2.1.5. Announcement 2.1.6. Subscription 2.1.7. Fetch Request 2.1.8. Relay 2.2. Model Content Protocol Concepts 2.2.1. MCP Server 2.2.2. MCP Client 2.2.3. Resources 2.2.4. Tools 2.2.5. Prompts 2.2.6. Request-Response 2.2.7. Notification 2.3. Verification Levels 2.3.1. Level 3: Official Pronouncements 2.3.2. Level 2: Corporate and Industry Sources 2.3.3. Level 1: Individual Comments 2.3.4. Level 0: Unverified Content 2.4. Protocol Architecture 3. MCP Endpoints Interactions 3.1. MCP Messages 3.1.1. REQUEST Structure 3.1.2. RESULT Structure 3.1.3. ERROR Structure 3.1.4. RESOURCE 3.1.5. TOOL 3.1.6. PROMPT Structure 3.1.7. Concept Mapping 3.2. Transport Mechanisms 3.2.1. Set up 3.2.2. Announce 3.2.3. Subscribe 3.2.4. Fetch 3.2.5. Terminate 4. Implementation Guidelines 4.1. Best Practices 4.2. Performance Considerations 5. IANA Considerations 6. Security Considerations 7. References 7.1. Normative References 7.2. Informative References 8. Acknowledgments Author's Address 1. Introduction 1.1. Motivation 1.1.1. Leveraging AI AI Agents have the potential to revolutionize various domains by automating and enhancing data exchange, decision-making, and interactions with external systems. However, their widespread adoption depends on ensuring that they are equipped with accurate, accessible, and verifiable data. This protocol aims to provide the framework that AI agents can use to interact with data sources effectively while ensuring minimal risk of misinformation and guaranteeing that AI-driven outputs are reliable. 1.1.2. Factual Integrity To facilitate the effective use of AI-generated content by humans, it is crucial that the generated information remains highly faithful to factual accuracy. At a minimum, AI agents must ensure that referenced information is neither distorted nor misrepresented, and that the generated content maintains logical consistency. Furthermore, factual integrity is essential for building trust in AI- generated outputs. When AI systems produce content that aligns with verifiable sources without introducing misinformation, users can rely on them with greater confidence. Maintaining factual integrity also reduces the risk of spreading inaccuracies, thereby enhancing the reliability and usability of AI-generated content across various domains. 1.1.3. Reachability AI agents are required to generate outputs based on various inputs, including resources that may not be explicitly specified. Therefore, it is essential that all necessary resources are accessible to AI agents and that they are aware of where to obtain specific information. This ensures that AI agents can function effectively, without limitations, in retrieving and processing the data required for accurate outputs. 1.1.4. Accessibility Web resources may restrict AI agents from scraping content, particularly for critical components such as advertisements or other important sections of a site. This restriction could result in the inaccessibility of essential data, limiting the AI's ability to function optimally. The protocol addresses these challenges by outlining strategies for AI agents to respect data access rules while still ensuring valuable resources are reachable when needed. 1.1.5. Traceability To ensure the reliability and transparency of AI-generated outputs, it is necessary to establish traceability mechanisms. These mechanisms allow for tracking the sources and reasoning behind the data used by AI agents in generating content. This will help verify the quality of the data and provide accountability for the outputs generated, which is crucial in sensitive contexts such as legal, medical, and financial applications. 1.1.6. Latency Reducing latency in communication between AI agents and data sources is vital for ensuring responsive and real-time interactions. The protocol establishes guidelines for minimizing latency and ensuring that AI agents can efficiently process and respond to data requests in a timely manner. Optimized transport mechanisms will contribute to lowering delays and improving the overall user experience. 1.2. Purpose 1.2.1. Large Integrated Resource Establish a robust network of Model Content Protocol (MCP) servers to facilitate seamless communication between AI agents and various external data sources. This network will serve as the backbone for enabling efficient data retrieval, processing, and integration into AI systems, ensuring smooth and consistent interactions across different platforms. 1.2.2. Educational Use of AI Promote the greater adoption of AI in educational contexts, enhancing learning methods through AI agents that interact with students, provide personalized feedback, and facilitate dynamic learning experiences. AI agents can help create customized educational environments, where students’ progress is tracked and tailored learning paths are provided based on their individual needs. 1.3. Scope The scope of this document includes the design and specification of a transport protocol that enables AI agents to interact with web resources and external data sources effectively. This protocol ensures that AI agents maintain factual integrity, adhere to access policies, and minimize latency, all while supporting educational and other use cases. It covers the communication mechanisms, data formats, security considerations, and best practices for implementing and utilizing this protocol. 1.4. Terms and Definitions * *AI Agent*: A system that uses artificial intelligence to process data, generate content, or interact with users and other systems. * *MCP Server*: A server that implements the Model Content Protocol to facilitate the integration of external data sources with AI applications. * *Web Resource*: A source of data available on the web, which may include web pages, databases, APIs, or other online content. * *Transport Protocol*: A protocol designed to facilitate the exchange of data between systems, ensuring efficient and secure communication. * *Factual Integrity*: The quality of information being true, accurate, and free from distortion or misrepresentation. 2. Concepts 2.1. Media over QUIC's Concepts Media over QUIC (MOQ) is a live media protocol powered by QUIC 2.1.1. Transport In this protocol, use the MOQTransport protocol for transport, ensuring the secure and efficient transmission of data between AI agents and external systems. 2.1.2. Session The session of the MOQ concept. The MOQ session provides a framework for managing ongoing interactions between AI agents and external data sources, ensuring that context and state are preserved during the communication process. 2.1.3. Track A single resource delivered as a "track" within the MOQ system. A track may refer to any distinct data set, file, or content that is used in the protocol’s operations. 2.1.4. Group A group is a set of data sent in a single QUIC stream. 2.1.5. Announcement The MOQ's Announcement. It indicates where endpoints can find resources such as MCP server. 2.1.6. Subscription The MOQ's Subscription. A single subscription is a single resource. 2.1.7. Fetch Request The MOQ's Fetch Request. Endpoints send request to the other as the MOQ's Fetch Request. Thus request and responce will be negotiated on a single QUIC stream. 2.1.8. Relay The MOQ relay is a component that facilitates the transfer of data between AI agents and external resources, ensuring that communication is efficient and reliable. 2.2. Model Content Protocol Concepts Model Content Protocol (MCP) is an open protocol designed to enable seamless integration between large language model (LLM) applications and external data sources. It facilitates secure data exchanges, improving the accessibility and utility of AI systems. 2.2.1. MCP Server MCP servers provides information from resources to MCP clients 2.2.2. MCP Client MCP clients comminucate with MCP servers and accesses to modified resources through the server. 2.2.3. Resources The Resources of the MCP concept. Every resources maps to a single track. * *Remote Resource*: A resource that is served from a remote server, such as an API response. * *Local Resource*: A resource that is located within a local environment, such as a database or file system, providing content that is directiory accessible by the AI agent. 2.2.4. Tools The Tools of the MCP concept. Functions that can be called by the LLM 2.2.5. Prompts The Prompts of the MCP concept. Pre-written templates that help users accomplish specific tasks 2.2.6. Request-Response An endpoint sends a REQUEST message to another endpoint. The receiving endpoint verifies if it has a method that matches the Method Name and processes the Parameters. The endpoint that received the request responds by sending a RESULT message. 2.2.7. Notification An endpoint informs other endpoints about the methods it supports by sending a NOTIFICATION message. 2.3. Verification Levels Verification levels indicate the reliability of the generated data. 2.3.1. Level 3: Official Pronouncements Content generated from authoritative sources, such as government reports or officially published information, which carries a high level of reliability. 2.3.2. Level 2: Corporate and Industry Sources Content generated from corporate websites, industry reports, and professional publications. These sources are generally reliable but should still be verified. 2.3.3. Level 1: Individual Comments Content generated from individual comments on the web, such as social media posts, blog entries, and forum discussions. These sources vary in reliability and should be verified for accuracy. 2.3.4. Level 0: Unverified Content Content from anonymous sources or user-generated content that has not been verified. This level carries the lowest reliability. 2.4. Protocol Architecture 3. MCP Endpoints Interactions 3.1. MCP Messages MCP messages are JSON-RPC 2.0 style. 3.1.1. REQUEST Structure REQUEST { id: Request ID (i), method: Method Name (b), parameters: [ { id: Parameter ID (i), value: Parameter Value (b), } ] } 3.1.2. RESULT Structure RESULT { id: Request ID (i), [key: Responce Key (b)]: Responce Value (b), } 3.1.3. ERROR Structure ERROR { code: Error Code (i), reason: Reason (b), data: Data (b), } ### NOTIFICATION Structure ~~~text NOTIFICATION { id: Notification ID (i), method: Method Name (b), arguments: [ { id: Argument ID (i), type: Argument Type (i), } ] } 3.1.4. RESOURCE 3.1.5. TOOL 3.1.6. PROMPT Structure PROMPT { name: Prompt Name (b), parameters?: [ { id: Parameter ID (i), name: Parameter Name (b), required: Required (i), }... ] } 3.1.7. Concept Mapping * "prompts/list" endpoint -> A Track named "prompts" 3.2. Transport Mechanisms 3.2.1. Set up A session is established by a client connecting to a server using the MOQ (Media over QUIC) protocol. During the setup phase, the client initiates a connection request to the server. Any additional information required for the session, such as authentication tokens or configuration parameters, should be included as extensions to the MOQ setup messages. This ensures that both the client and server have all necessary information to establish a secure and efficient communication channel. 3.2.2. Announce Endpoints can discover available resources through MOQ announce negotiation. This process begins with a subscriber indicating to a publisher the specific resources they are interested in. The publisher then responds with an announcement detailing the available resources. This negotiation allows for dynamic resource discovery and ensures that subscribers are aware of the resources they can access. The server's announcement includes metadata about the resources, such as their type, availability, and any access restrictions. 3.2.3. Subscribe Reserved Tracks 3.2.4. Fetch 3.2.5. Terminate 4. Implementation Guidelines 4.1. Best Practices 4.2. Performance Considerations 5. IANA Considerations 6. Security Considerations 7. References 7.1. Normative References 7.2. Informative References 8. Acknowledgments Author's Address Okutani Daichi FOALK Email: daichi1616.kytuniv@gmail.com