What is the Agentic AI Foundation (AIF) and what is its purpose?

The AIF is a new open-source foundation launched by Anthropic, OpenAI, and Block, under the Linux Foundation. Its primary purpose is to provide a neutral home for agentic AI projects, ensure open standards, foster industry collaboration, and support the development of a robust, interoperable agentic ecosystem.

How does MCP handle authentication, especially for enterprises?

MCP initially combined authentication and resource servers, which proved problematic for enterprises using central identity providers (IDPs). A June spec update separated these, making the MCP server a resource server while providing guidelines for obtaining tokens from separate authentication servers, thus enabling better enterprise integration with existing IDPs.

What were the challenges with MCP's transport protocol and HTTP streaming?

The design of the transport protocol aimed for both simplicity (REST-like) and full bidirectional streaming (like WebSockets). However, making client return streams optional led to a loss of bidirectionality and features like elicitations. Additionally, holding state on the server side created scalability issues for large-scale deployments, requiring new solutions for shared state management.

How do model improvements influence MCP protocol design?

While models are not directly involved in MCP, model improvements influence its design by anticipating capabilities like progressive discovery. The protocol relies on the models' ability to understand tools implicitly and progressively discover them, allowing for a leaner context window rather than bloating it with all available tools upfront. Training models for this capability further optimizes their performance.

What is the difference between MCP and 'skills' in the context of AI agents?

MCP and skills are largely orthogonal. Skills provide domain knowledge and specific task behaviors (e.g., 'act as a data scientist'), often requiring an execution environment. MCP, on the other hand, provides the communication layer for connecting agents to external actions and data sources, including authentication and external services. They can be used together, with skills using MCP servers to access company data.

What are 'long-running tasks' in MCP and why were they introduced?

Long-running tasks were introduced as a native primitive in MCP to support operations that might take hours or even days, like deep research or complex agentic workflows. Previously, users awkwardly tried to manage these with standard tools, which lacked proper support for intermediate results and asynchronous polling. Tasks provide a first-class way to handle these persistent, asynchronous operations.

How will MCP Apps (UI components via MCP) enhance AI applications?

MCP Apps allow AI applications to embed richer, visual user interfaces by serving raw HTML into iframes. This addresses the limitations of purely text-based interfaces for tasks like seat selection, shopping, or music production where visual interaction is superior. It aims to create a common standard for building interactive UIs that models and humans can both interact with seamlessly.

What is Block's Goose and its role within the Agentic AI Foundation?

Goose, developed by Block, is an open-source agent interface and coding tool. It was an early adopter and contributor to MCP and has been donated to the AIF. Goose serves as a concrete reference implementation, demonstrating how MCP and new agentic standards like MCP Apps can be used in real-world scenarios, extending beyond just coding to various other applications.

What are the key benefits for a project to join the Agentic AI Foundation?

Joining the AIF provides a neutral home for a project, ensuring its openness and long-term viability. It fosters industry-wide collaboration, leading to better support, increased development velocity, and higher code quality through collective investment. For protocols and standards, it offers wider adoption and a common forum for technical communication and robust development.

Why was a new foundation built for agents rather than joining existing Linux Foundation initiatives like Data and AI?

The Agentic AI Foundation was established because agents represent a distinct enough set of technology and audience compared to other AI areas like foundational models or inference infrastructure (e.g., PyTorch, Kubernetes). This dedicated focus allows for effective community building, curation, and tailored support for the specific challenges and innovations within the agentic systems space.

What are the actionable outcomes and investments of the Agentic AI Foundation?

The AIF will manage intellectual property, facilitate leadership and consensus building, and fund developer outreach and community engagement. It will also host events (e.g., conferences, virtual meetings) to bring developers together for co-development and decision-making. Resources will also be allocated to build platforms for community metrics, security practices, and understanding project adoption to foster a virtuous cycle of improvement and commercial success.

One Year of MCP — with David Soria Parria and AAIF leads from OpenAI, Goose, Linux Foundation

Latent Space Podcast

People & Blogs6 min read100 min video

Dec 28, 2025|4,388 views|83|5

Save to Pod

Key Moments

On this page

TL;DR

MCP protocol is one year old, donated to the Agentic AI Foundation for neutral governance and continued growth.

Key Insights

The Model Context Protocol (MCP) has seen rapid adoption in its first year, with key inflection points including endorsements from major tech leaders and companies.

MCP has undergone significant protocol development, moving from local-only to supporting remote servers with enhanced authentication and introducing long-running tasks.

The Agentic AI Foundation (AAIF) has been launched to provide a neutral governance umbrella for MCP and other complementary protocols, prioritizing adoption and neutrality.

The development of MCP has benefited from industry collaboration, including input from experts on OAuth, and a focus on security best practices.

MCP's future development aims to improve scalability, support agent-to-agent communication, and make the protocol more stable after establishing solid foundations.

The AAIF aims to foster a collaborative ecosystem for agentic AI, enabling developers to build interconnected AI applications and tools.

MCP'S RAPID ADOPTION AND EVOLUTION

The Model Context Protocol (MCP) marks its first public anniversary, having experienced significant growth and adoption in just one year. Initial adoption by builders and early clients like Cursor and VS Code rapidly accelerated, reaching a major inflection point around April with endorsements from industry leaders at Microsoft, Google, and OpenAI. This widespread adoption spurred continuous development on the protocol itself, including the introduction of remote server support, robust authentication mechanisms, and long-running task capabilities, setting the stage for future scalability and agent-to-agent communication.

THE FORMATION OF THE AGENTIC AI FOUNDATION (AAIF)

To ensure MCP's future as a neutral and open standard, Anthropic, alongside collaborators like OpenAI and the Linux Foundation, has launched the Agentic AI Foundation (AAIF). This foundation will serve as an umbrella organization, safeguarding the neutrality of projects under its governance. The goal is to provide a secure and trusted environment for the development and adoption of nascent AI protocols, ensuring they remain accessible and beneficial to the broader industry.

PROTOCOL DEVELOPMENT AND ENHANCEMENTS

MCP's first year saw several key protocol releases, addressing critical needs like HTTP streaming and sophisticated authentication. The March spec introduced remote server support and initial authentication, which was refined in the June spec with the help of industry experts in OAuth. The focus on security best practices and enterprise needs has been paramount. Most recently, the November release introduced long-running tasks, paving the way for more complex agentic operations and inter-agent communication.

TRANSPORT AND SCALABILITY CHALLENGES

Designing a transport protocol that balances simplicity with the need for full bidirectional streaming has been a significant challenge for MCP. While streamable HTTP was chosen for its reliance on standard HTTP, making bidirectional features optional led to reduced functionality. Scalability, particularly for horizontally scaled servers, also presents complexities in managing state. Current efforts are focused on refining the protocol to be both simple for basic use cases and robust for complex, scalable agent interactions.

COLLABORATION AND STANDARD SETTING

The development of MCP and now the AAIF involves intense collaboration among major tech players like Google, Microsoft, OpenAI, and Anthropic. These collaborations, often resembling IETF processes but with increased speed, focus on defining core problems and building solutions. The goal is to create a standardized communication layer for AI applications, avoiding fragmentation and fostering a cohesive ecosystem. This cooperative effort is crucial for navigating the rapid pace of AI development.

MCP'S ROLE IN THE DEVELOPING AGENTIC ECOSYSTEM

MCP is positioned as a communication and connectivity layer, distinct from 'skills' which provide domain-specific knowledge. While skills enable models to perform specialized tasks, MCP facilitates the actual actions and data interactions with the outside world, including critical authentication. It aims to simplify AI application development by providing a standardized way to connect models to external tools and data, making it easier for developers to build sophisticated AI agents.

INTERNAL ADOPTION AND INFRASTRUCTURE

Major companies, including Anthropic, heavily 'dogfood' MCP internally, building custom MCP servers and gateways to integrate with internal tools and services. This internal usage demonstrates the protocol's value for enterprise applications, from summarizing Slack messages to conducting company-wide surveys. The development of user-friendly infrastructure, such as one-command deployments for MCP servers, is key to facilitating widespread adoption, both internally and externally.

REGISTRIES AND DISCOVERY MECHANISMS

The AAIF aims to establish a centralized registry system, similar to npm, for MCP servers to facilitate discovery and adoption. While public registries offer broad access, they also pose risks like supply chain attacks. The vision includes multi-layered registries, where companies can maintain curated internal registries reflecting their trust standards. The goal is to enable models to intelligently discover and utilize appropriate MCP servers for specific tasks, moving towards a more seamless AI interaction model.

EVOLVING USE CASES AND NEW FEATURES

MCP is primarily used for data consumption and providing context to AI models. However, the introduction of long-running tasks signifies a move towards supporting more complex workflows and agentic operations. New features like MCP Apps are also emerging, focusing on richer, visual interfaces beyond text-based interactions, which are crucial for domains like flight booking or UI-driven tasks. The protocol's evolution is driven by user needs and emerging AI capabilities.

LONG-RUNNING TASKS AND ASYNCHRONOUS OPERATIONS

The newly introduced 'task' primitive in MCP addresses the critical need for long-running operations, essential for advanced agents and deep research tasks. Unlike simple RPC calls, tasks are designed to handle operations that may take hours or days, potentially returning intermediate results and supporting asynchronous execution. This feature aims to standardize how agents perform extended operations, moving beyond the limitations of synchronous tool calls and providing a more robust framework for complex AI workflows.

MCP VERSUS REST APIS AND CONSUMER FOCUS

While MCP is often compared to REST APIs, it offers distinct advantages, particularly in handling authentication and managing tool bloat when interacting with AI models. MCP is envisioned as a consumer-focused protocol, where end-users interact with AI applications without needing to understand the underlying MCP layer. Its primary value lies in enabling seamless pluggability of external services into AI applications, abstracting away the complexities for the end-user.

MCP APPS AND RICH USER INTERFACES

MCP Apps represent an extension to MCP focused on enabling rich, interactive user interfaces for AI applications. This is crucial because pure text-based interactions are limiting for many tasks. By allowing models to navigate and interact with visual elements, MCP Apps aim to provide more natural and effective user experiences, particularly in areas like shopping or seat selection where visual representation is key. These extensions are built using standard HTML within iframes, facilitating rich interactions.

BROADER APPLICATIONS AND EXTENSIONS

Beyond core communication, MCP is being extended to address industry-specific needs, such as financial services and healthcare. These extensions involve implementing specific security guarantees, attribution requirements, and data handling policies, like HIPAA compliance. Such extensions ensure that MCP can support sensitive data interactions and adhere to legal and regulatory mandates, expanding its applicability beyond general-purpose AI tools.

THE FUTURE OF AGENTIC AI AND THE FOUNDATION

The launch of the AAIF signifies a commitment to fostering a collaborative and open ecosystem for agentic AI. The foundation plans to provide resources for community building, developer outreach, and event hosting. Future work will focus on enabling asynchronous agent interactions, facilitating deeper AI integration into business processes, and ensuring that MCP and related technologies continue to evolve to meet the demands of increasingly sophisticated AI applications.

Mentioned in This Episode

●Tools & Products

●People Referenced

MCP Development and Foundation Participation

Practical takeaways from this episode

Do This

Actively build good MCP servers for diverse use cases.

Engage with the MCP community through discord, feedback, and ideas.

Contribute to SDK development (TypeScript, Python, Go, Haskell, etc.) for MCP.

Utilize new MCP features like long-running tasks for complex agent operations.

Consider using the MCP framework for secure external system integration, especially in regulated industries.

Embrace private, curated registries for trusted internal MCP server discovery within companies.

Focus on composable, interoperable projects within the Agentic AI Foundation (AIF).

Avoid This

Don't build only naive, simple MCP servers; explore complex actions and workflows.

Don't ignore authentication for remote MCP servers, ensure proper OAuth implementation.

Don't combine authentication and resource servers in a single MCP server for enterprise use.

Don't overlook the importance of bidirectional streaming in MCP transport for advanced features (e.g., elicitations, sampling).

Don't treat skills and MCP as overlapping; understand their orthogonal roles for domain knowledge versus external connectivity.

Don't assume public registries are fully trusted; be aware of supply chain attack risks.

Don't expect the Agentic AI Foundation to push speculative architectures; focus on established, well-maintained projects.

Common Questions

MCP (Model Context Protocol) is an open-source communication layer between AI applications and servers. Over the past year, it has seen massive adoption, including by major clients like Cursor and VS Code, and by large companies like Microsoft, Google, and OpenAI. Key evolutions include moving from local-only to remote MCP servers, significant improvements to the authentication spec, and the introduction of long-running tasks for complex operations.

Topics

MCP Protocol Agentic AI Foundation (AIF)HTTP Streaming Protocol Design Progressive Discovery Skills Vs. MCP Model Context Long-Running Tasks MCP Apps UI For AI Linux Foundation Community Building Goose Context Compression

Mentioned in this video

toolSmitheries

Mentioned as an example of sub-registries that can filter and curate MCP servers.

toolHaiku model

A type of small model, mentioned hypothetically for context compression to determine what information an agent needs to retain during long-running tasks.

personErs Holtzel

Mentioned in the context of Google's decision to open-source Kubernetes to the CNCF.

toolAgent AI Foundation (AIF)

A newly launched open-source foundation, formed by Anthropic, OpenAI, and Block, under the Linux Foundation, to standardize agentic AI development.

toolFast MCP Cloud

A company/service offering simple deployment of MCP servers with two commands.

personCVS Pharmacy

Mentioned as an example of a 'late majority' user and adopter of technology, indicating widespread industry adoption.

toolOAuth

An open standard for access delegation, used in MCP for authentication, with specific challenges in enterprise contexts.

personBrian Stevens

Mentioned in the context of Google's decision to open-source Kubernetes to the CNCF.

personJohn Welsh

Mentioned as giving a talk about Anthropic's internal MCP gateway usage.

toolSlack MCP server

An internal MCP server at Anthropic used to summarize Slack conversations.

toolGithub registry

A curated registry for MCP servers that speaks the same format as the official MCP registry.

toolTurkish Airlines MCP server

A specific MCP server built by Turkish Airlines to search for flights and related services.

toolGoose

An open-source agent interface and coding tool, donated to the Agentic AI Foundation, which was an early adopter and contributor to MCP.

personNick Cooper

From OpenAI, described as the head of protocol things and a core contributor to MCP, representing OpenAI for the AIF.

toolMCPUI / MCPI

An extension to MCP that allows for exposing UI components, serving raw HTML over an MCP resource into an iframe for richer interactive interfaces for AI applications.

personJim Zlin

CEO of the Linux Foundation for 22 years, who helped facilitate the launch of the Agentic AI Foundation.

toolLinux Foundation

The largest open-source foundation, providing a neutral home and organizational support for projects like MCP and the Agentic AI Foundation.

toolagents MD

An OpenAI contribution to the Agentic AI Foundation.