Key Takeaways
- Anthropic's Model Context Protocol (MCP) has become a de facto standard for agentic systems, adopted widely by major tech companies.
- The Agentic AI Foundation (AAIF) has launched under the Linux Foundation, with Anthropic, OpenAI, and Block as founding members.
- MCP's protocol has evolved to support HTTP streaming and enterprise-grade authentication, separating resource and identity servers.
- Enterprises are rapidly adopting MCP internally, utilizing gateways for secure integration with proprietary data and compliance-heavy workflows.
- New MCP features include 'tasks' for long-running asynchronous agent operations and 'MCP Apps' for richer, interactive user interfaces.
- The AAIF aims to curate valuable open projects based on principles of composability, technical merit, and demonstrated traction.
- The Linux Foundation provides a neutral home, investing in community growth, project quality, and commercial adoption for AAIF initiatives.
Deep Dive
- Collaboration with senior engineers from Google, Microsoft, AWS, Anthropic, and OpenAI was key to reaching consensus on MCP protocol design.
- The standardization process is compared to IETF, but MCP's private working group aims for greater efficiency and faster development than multi-year cycles seen in standards like OAuth 2.1.
- MCP's project structure balances consensus with a core maintainer group to facilitate rapid development, described as a "moderated dictatorship."
- Model improvements, especially in tool calling and post-training, directly influence MCP's protocol design, with a focus on anticipating advancements.
- Anthropic extensively uses MCP internally, including an MCP gateway for security and custom servers for Slack summarization and employee surveys.
- Enterprises are adopting MCP rapidly by using gateways that simplify authentication and protocol details, allowing developers to run MCP servers with standard I/O.
- This widespread internal adoption, often unnoticed, is happening at massive scale in large companies and stems from a need to scale development tooling faster than company growth.
- The need for a centralized, 'npm-like' registry for agents is highlighted, alongside existing options like Smithery, for simplifying discovery and installation.
- Discussions included distributed code signing for MCP tool descriptions, with Anthropic and OpenAI providing signatures to establish trust.
- Private registries within companies offer a secure environment for MCP servers, supporting compliance with regulations like HIPAA and financial data privacy, crucial for use cases such as financial services and Turkish Airlines' flight search.
- MCP's 'tasks' feature was developed for long-running, asynchronous agent operations, addressing user demand beyond simple tool calls.
- Tasks enable complex research and agent handoffs, with the goal of supporting operations lasting hours or even days.
- Design decisions focused on creating a generic 'container' for asynchronous operations, allowing for future extensibility beyond just tool calls, and supporting both polling and eventual webhook notifications for completion.
- MCP aims to be an invisible communication layer for AI applications, enabling richer, visual interfaces beyond text, using iframes for MCP Apps.
- These apps allow direct interaction with UI elements like seat selection for flights or shopping interfaces, mirroring trends seen with ChatGPT apps.
- Collaboration with OpenAI is ongoing to build a common standard for MCP Apps, addressing potential cross-origin issues and the need for consistent branding and styles.
- David Soria Parra affirmed Anthropic's strong commitment to MCP, with the same team members and product focus maintained post-foundation launch.
- The Agentic AI Foundation ensures MCP's perpetual openness and neutrality, establishing a precedent for open standards in the industry.
- The AAIF core team includes Jim Zemlin (Linux Foundation CEO), Nick Cooper (OpenAI), David Soria Parra (Anthropic), and Brad Howes (Block).
- The Agentic AI Foundation (AAIF) has been formed under the Linux Foundation, with Anthropic, OpenAI, and Block as founding members.
- The foundation hosts Anthropic's Model Context Protocol (MCP) and Block's Goose coding agent, aiming to be a neutral home for agentic systems.
- Jim Zemlin will serve on the AAIF's technical steering committee, tasked with curating important projects and preventing the foundation from becoming a "dumping ground."
- Brad Howes from Block discussed Goose's involvement with the AAIF, emphasizing the need for concrete, open-source reference implementations.
- Goose is integrating MCP Apps, a new specification for richer UI elements, to showcase abstract protocol features through practical applications like dashboards and enhanced chat experiences.
- Goose was an early adopter and collaborator, integrating MCP before its official launch and becoming one of the first non-Anthropic contributors, demonstrating the protocol's utility.
- The foundation does not have a fixed roadmap but operates on principles for selecting projects, prioritizing composability for creating interoperable agentic systems.
- A 'tastemaking' curation process, led by elite architects and developers, is favored over vendor-driven decisions to ensure technical merit and relevance for foundation projects.
- A core principle is to accept projects that have already gained traction, are well-maintained, and are healthy, with contributions enhancing protocol robustness and implementation quality.
- Jim Zemlin explains that the Linux Foundation supports open-source projects by investing in platforms that help maintainers track community velocity, developer growth, and adoption indicators.
- A 'directed fund' within the foundation means stakeholders allocate money for specific market and community building, separating technical governance from financial spending.
- Zemlin highlights that the AAIF launch is experiencing significantly faster momentum than the Cloud Native Computing Foundation (CNCF) did, benefiting from 10 years of lessons learned.