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

David Soria Parra recaps MCP's explosive growth from Thanksgiving launch to industry-wide adoption by Microsoft, Google, and OpenAI. The protocol moved from local-only to remote servers with authentication, culminating in the Agentic AI Foundation launch under Linux Foundation with 50+ founding members.

• •Major inflection point in April when Sam Altman, Satya Nadella, and Sundar Pichai publicly committed to MCP adoption
• •Four major spec releases in one year: local servers, remote HTTP streaming (March), improved OAuth (June), and long-running tasks (November)
• •Foundation formed with Anthropic, OpenAI, and Block as founding projects, with industry-leading experts from AWS, Microsoft, Google participating in protocol development
• •MCP now has massive enterprise adoption with millions of requests at scale, though mostly internal/private deployments

Deep dive into MCP's OAuth implementation challenges and the critical June revision that separated authentication servers from resource servers. The team worked with OAuth spec authors to get enterprise authentication right, learning hard lessons about combining auth components.

• •Initial March spec incorrectly combined authentication server and resource server into MCP server, making enterprise IDP integration impossible
• •June revision separated concerns: MCP server is resource server, authentication happens via separate IDP (Okta, etc.)
• •Worked directly with OAuth specification authors to push boundaries of what OAuth can do for AI applications
• •Financial services and healthcare require special authentication extensions for legal compliance (attribution, HIPAA data isolation)

Discussion of the transition from local Standard IO to HTTP streaming transport, and the ongoing challenges of building a protocol that works for both simple tool providers and complex bidirectional agent communication at massive scale.

• •Streamable HTTP designed to retain stateful properties of Standard IO while enabling remote deployment
• •Made too many features optional for clients, resulting in no client implementing bidirectional streaming features like sampling
• •Current design requires server-side state management, creating scaling issues for companies doing millions of requests
• •Two-day workshop with Google, Microsoft, AWS engineers working on next iteration to solve stateless scaling while preserving agent capabilities

Introduction of the Tasks primitive for long-running operations like deep research agents. Tasks enable asynchronous execution with intermediate results, solving a major limitation where people were awkwardly trying to implement multi-hour operations using synchronous tools.

• •Tasks designed as containers for asynchronous operations, not just async tool calls - enables exposing full coding agents as MCP servers
• •Can return intermediate results and execution traces, showing which tools were called and what inputs were used
• •Polling-based with planned webhook optimization for completion notifications (similar to OS kernel patterns like inotify)
• •Addresses top enterprise request for operations that may not complete for hours or days, enabling true agent-to-agent communication

Clarification that 'code mode' (Cloudflare's term for programmatic MCP) is not replacing MCP but rather an optimization layer. MCP still provides authentication, discoverability, and self-documentation while models can compose multiple tool calls into executable code for efficiency.

• •Code mode is token optimization - model generates code that calls multiple MCP tools instead of sequential tool calls
• •MCP's value (authentication, discoverability, self-documentation) remains regardless of execution pattern
• •Progressive discovery allows models to explore available tools incrementally rather than loading everything into context
• •Models can already do progressive discovery without special training; post-training just optimizes the behavior

Skills provide vertical domain knowledge (how to behave as data scientist, accountant, etc.) while MCP provides horizontal connectivity to external systems. They're orthogonal and work best together, with skills using MCP servers for actual data access and actions.

• •Skills give domain knowledge and behavior patterns; MCP gives connectivity to external actions and data
• •Skills require execution environment and lack authentication; MCP handles both
• •Best deployment pattern: use skills to define different roles/functions, use MCP servers to connect those skills to company data sources
• •Skills are fixed in space (code/scripts), MCP servers can be externally maintained and improved without updating skills

Anthropic extensively dogfoods MCP internally with a custom gateway for easy deployment. Employees build their own MCP servers for everything from Slack summarization to survey analysis, with most servers unknown to the core team - validating the original vision of democratized tool building.

• •MCP gateway allows one-command deployment of servers with automatic authentication/IDP integration
• •Internal servers include Slack summarization, survey analysis, and many research-specific tools David doesn't even know about
• •MCP born from need to scale developer tooling faster than team could grow - let everyone build their own integrations
• •Similar platforms exist commercially (FastMCP Cloud, LightLLM gateway) for enterprises to replicate this pattern

Official MCP registry designed as NPM/PyPI equivalent with standardized API, enabling curated sub-registries (like Smithery) and private enterprise registries. Goal is model-driven auto-discovery and installation, but requires trust levels and potentially distributed signatures from model providers.

• •Official registry has same supply chain security issues as NPM - anyone can publish, need curation layer
• •Sub-registries (Smithery, GitHub, internal enterprise) speak same API format for interoperability
• •Vision: models auto-select and install MCP servers from trusted registries based on task requirements
• •May need distributed signing system where Anthropic, OpenAI, etc. sign trusted servers, or enterprise private registries with implicit trust

MCP Apps extension enables servers to provide rich HTML interfaces for tasks poorly suited to text (seat selection, shopping, music production). Implemented as iframes with post-message communication, with OpenAI and Anthropic working toward common standard after separate initial implementations.

• •Pure text interfaces fail for visual tasks like seat selection, shopping, theater booking - need application UI
• •Servers provide raw HTML in iframe, communicate via post-messages for security isolation
• •No external references to avoid CORS issues; styling inheritance from parent app still being worked out
• •Anthropic and OpenAI collaborating on unified standard after initial separate implementations (MCP UI vs ChatGPT apps)

The Agentic AI Foundation launched with 50+ companies and immediate overwhelming interest. Governance focuses on demonstrated utility over speculative standards, with technical steering committee using taste-making to curate composable, well-maintained projects that complement each other.

• •Selection principles: demonstrated traction, well-maintained, healthy community, complementary to existing projects
• •Protocols must show de facto standard status; applications can be more flexible but need proven utility
• •Composability is key - projects should work together like Lego blocks, not compete for same function
• •Technical steering committee (not vendor committee) makes curation decisions to avoid 'dumping ground' problem other foundations face

Goose serves as concrete reference implementation showing what MCP enables beyond abstract specs. As first non-Anthropic MCP contributor (day 2), Goose validates protocol changes and demonstrates real-world usage patterns from coding to science experiments to Google Docs.

• •Goose was first open source agent to integrate MCP, first external contributor to protocol
• •Provides concrete validation of abstract features like MCP Apps - shows actual user value
• •Can disable coding features entirely and use as general-purpose MCP client for any domain
• •Day-one partner for MCP UI development, helping prove out new protocol extensions before wider release

MCP Dev Summits in San Francisco and London revealed unexpected enterprise requirements, especially from financial services. Legal contracts require data attribution, regulatory compliance needs like HIPAA data isolation, and other constraints that don't exist in normal development.

• •Financial services have legal requirements to attribute third-party data sources in all displays
• •Bloomberg leading financial services working group to define client requirements for regulated MCP servers
• •Healthcare/HIPAA data may require servers to enforce that clients disable other MCP servers during session
• •These domain-specific extensions will likely become certification programs (financial services MCP client certification, etc.)

Moving to Linux Foundation ensures MCP stays permanently open and neutral, preventing proprietary capture like HDMI 2.1. Provides industry confidence for long-term investment while maintaining Anthropic's technical commitment. Foundation enables community building beyond what single company can achieve.

• •Prevents proprietary capture - HDMI Forum blocking open source Linux drivers is cautionary tale
• •Gives industry confidence to invest in implementations knowing protocol stays open and neutral
• •Anthropic commitment unchanged - David still lead maintainer, same team, same product integration
• •Foundation provides venue for industry collaboration that single company can't offer, even with open source license