Latent Space: The AI Engineer Podcast

Episode: One Year of MCP — with David Soria Parra and AAIF Leads from OpenAI, Goose, Linux Foundation
Date: December 27, 2025

Overview

This episode celebrates the one-year anniversary of the Model Context Protocol (MCP) and delves into its explosive adoption, technical evolution, and formalization within the newly created Agentic AI Foundation (AAIF) under the Linux Foundation. Host Alessio and Squiggs are joined by David Soria Parra (Anthropic), Nick Cooper (OpenAI), Brad (Block/Goose), and Jim Zemlin (Linux Foundation) to break down the past year, key developments, lessons learned, and the significance of a neutral, open-source standard for agentic AI systems.

Key Discussion Points & Insights

MCP’s First Year: History & Growth

• Rapid Adoption:
  • Initial “crazy growth” right after launch, with early uptake from VSCode, Cursor, and community builders.
  • Adoption hit an inflection point in April 2025 when industry leaders (Sam Altman, Satya, Sundar) endorsed MCP for Microsoft, Google, and OpenAI.
  • “It felt like an eternity, honestly… but we have this crazy growth and adoption.” (David, 01:14)
• Technical Evolution and Milestones:
  • Spec releases: Four spec iterations addressing streaming, authentication, remote servers, and long-running tasks.
  • Security & Enterprise Concerns: Collaborated with OAuth experts to refine authentication for enterprise deployment.
  • Long-running Tasks: Introduced support for agent-to-agent and deep research tasks in latest protocol.
    • "...the next bigger improvement to the protocol which is long running tasks to really allow for deep research type of task and maybe even agent to agent communication." (David, 02:56)
• Origination & Motivation:
  • MCP was born out of a need at Anthropic for internal scaling and developer autonomy:
    • “How can I build something that [the teams] can all build for themselves?... and that's really the origin story of MCP.” (David, 28:11)

The Agentic AI Foundation (AAIF)

• Why a Foundation?
  • To guarantee lasting neutrality and openness of MCP, avoiding the pitfalls of protocols getting locked down by single entities.
  • “The foundation part is mostly just an umbrella to make sure the projects under it stay always neutral.” (David, 05:28)
• Role & Structure:
  • Overseen by a technical steering committee, driven by core maintainers—not a "dumping ground" but focused on well-maintained, impactful projects.
  • "We really want to make sure that we're not offering the foundation... five protocols for the same communication." (David, 04:52)
• Linux Foundation’s Perspective:
  • “AI years are kind of like dog years” (Jim, 68:02)
  • MCP’s velocity and rapidly proven adoption are rare among early-stage tech standards.
• Membership & Expansion:
  • Inaugural members include Anthropic, OpenAI, Block, major cloud and enterprise players, signaling cross-industry buy-in.

Protocol Deep Dives & Technical Reflections

Remote Servers & Transport Evolution

• From Local to Remote:
  • Initially supported local (on-prem) servers, now also remote, distributed models.
  • HTTP streaming became essential for remote, bi-directional communication.
• Authentication Lessons:
  • Early missteps in OAuth implementation—combined authentication and resource server, not enterprise friendly.
  • Fixed by separating these concerns, mirroring real-world IDP flows (Okta, Google).
    • “Most of them honestly were just me not understanding enterprises well enough.” (David, 08:51)
• Scalability & State Management:
  • Realized challenges with server-side state for horizontal scaling (k8s, multi-pod); technical efforts ongoing to make bidirectional and stateful communication scalable.
    • “That is fine if you have one server, but the moment you scale this horizontally... it becomes a problem.” (David, 14:25)

Model-Protocol Interplay

• Separation of Concerns:
  • MCP is a protocol between applications and servers, not the models; composability and context engineering matter most.
  • Protocol design anticipates, but is not rigidly adapted to, model advances.
  • “The primitives of the protocol, they're actually very rarely influenced by model improvements.” (David, 18:56)
• Handling Context Bloat:
  • Emphasizing progressive discovery rather than dumping all tools/context into the system prompt—“let the model decide to gain more information.” (David, 19:27)

Tasks, Skills, and New Application Patterns

• Skills vs MCP:
  • Skills: domain-specific behaviors, toolchains, or code scripts inside apps.
  • MCP: standardized secure connectivity for external actions and data.
    • “Skills are super great… but MCP gives you the connectiveness of the actual actions that you can take with the outside world.” (David, 23:42)
• Async Tasks:
  • MCP’s new task primitive enables long-running, asynchronous workflows—critical for complex agent orchestration.
    • “We design tasks because people come to us and go like, okay, we really want long running operation which is basically agents.” (David, 42:27)

Registries, Discovery & Ecosystem

• Official Registries:
  • Inspired by NPM/PyPI, but adapted for AI: aims for both open, uncurated registry and curated sub-registries for security/trust.
  • “The world we want to live in... a model can auto select an MCP server from a registry, install it, and then for the given task... use it.” (David, 31:20)
• Discovery Challenges:
  • AI-first context: models can intelligently discover and select servers/tools, unlike traditional package discovery.

MCP Apps & UI Standards

• Need for Visual Interactivity:
  • Text only UIs are limiting; new standard for serving HTML/iframes to create rich, interactive apps (e.g., flight seat selection).
  • Collaboration underway between Anthropic & OpenAI to standardize UI extensions.
    • “You build for one platform, but you can use it across all of them.” (David, 52:54)
• Technical Approach:
  • Use of HTML served over MCP resource, embedded as an iframe for isolation—but future improvements likely needed for branding and style pass-through.

Notable Quotes & Memorable Moments

• On Industry Collaboration:

  "Sometimes you step back and like you sit in a room with all these competitive companies, but you're actually building something together. And I love that."
  — David, 15:49

• On Open Standards:

  "The commitment of Anthropic is the same… But the main part of the foundation is making sure that the whole industry knows that this will stay forever open, that this cannot be taken away."
  — David, 61:17

• On Learning from Community:

  “The one thing, if anything, I'm proud of is like building a community of people that can come together and help me figure shit out, because I have my set of experiences...”
  — David, 08:31

• On Agentic Systems:

  “Agentic systems are a sum of many, many parts... interoperable composable bits that all work together.”
  — Nick (OpenAI), 78:23

• Linux Foundation on Direction & Resources:

  “We have a whole platform that enables maintainers to look at their community and understand, like, what's our velocity, how many developers are we adding?... how's our security doing?”
  — Jim, 89:45

Timestamps for Key Segments

| Segment | Timestamp | |---------------------------------------------|---------------| | MCP Past Year Overview & Inflection Points | 01:14–03:35 | | Protocol Design/Spec Releases | 06:41–11:42 | | Deep Dive: Streaming, Authentication, Scale | 11:42–18:30 | | Registries & Discovery | 30:42–36:16 | | Community & Events (Dev Summit, hackathons) | 36:16–40:30 | | Tasks / Long-Running Operations | 42:13–47:09 | | Model-Protocol Interactions | 18:30–23:12 | | UI Extensions & App Experience | 52:27–58:09 | | Foundation Launch & Governance | 63:28–66:42 | | Linux Foundation Vision, Incentives | 67:51–91:51 | | Looking Forward (Panel Responses) | 95:03–99:02 |

Community, Adoption, and Future Vision

• Internal Dogfooding:
  • Anthropic and others massively use MCP internally, from summarizing Slack to conducting internal surveys; many unknown servers run as employees “build for themselves.”
• Enterprise Uptake:
  • MCP’s most explosive adoption is inside large enterprises, often hidden from public view, mirroring early Docker/Kubernetes trajectories.
• Problems Learned from the Field:
  • Industry-specific requirements, especially in finance and healthcare, introduce new constraints on attribution, data access, and protocol extensions.
• Encouragement:
  • David urges engineers to “build really good MCP servers” and join the open community as contributors, testers, and SDK authors (59:44).

Conclusion: The Meaning of AAIF and the Road Ahead

• Neutral Ground for Agentic AI:
  • AAIF is set up as both a safe neutral zone for agentic AI protocols and a “town square” for coordination, with a focus on composability, adoption, and real-world utility.
• Founding Partners & Momentum:
  • Founding coalition includes Anthropic, OpenAI, Block, Linux Foundation, and over 50 companies from day one—unprecedented buy-in.
• Next Steps & Challenges:
  • The panel looks forward to richer asynchronicity for agents, reference implementations, asynchronous workflows, and deeper real-world adoption stories.
  • The open question: what technical contributions will define year two and beyond?

“This is really a beginning and so I want to see it be healthy and grow and I just don't know what comes next. So I'm most excited to see that.”
— Nick (OpenAI), 95:50

Panel Intros & Voices (63:43–64:53)

• Jim Zemlin (Linux Foundation): CEO, responsible for facilitating AAIF’s launch.
• Nick Cooper (OpenAI): Head of protocol initiatives, OpenAI’s rep for AAIF, MCP core contributor.
• David Soria Parra (Anthropic): Co-creator of MCP, technical leader at Anthropic.
• Brad (Block/Goose): Principal Engineer, open source lead, creator of Goose.

Call to Action

• Build: Try deploying and improving MCP servers—see use cases and share feedback.
• Contribute: Get involved in the open community, from code to standards development.
• Engage: Attend AGI-focused summits/hackathons and join the Discord/community channels.
• Watch the Foundation: Expect major leaps in composable agentic systems and real-world adoption coming soon.

For full show notes, head to latent.space.