110 Million SDK Downloads per Month

▶ Watch (0:34)

MCP reached 110 million SDK downloads per month within 16 months of launch. React took roughly 3 years to hit that download volume. The SDK dependencies include OpenAI Agents SDK, LangChain, and Pydantic AI. Bishop framed the adoption as evidence that the ecosystem needed a standard to connect tools to AI systems. Before MCP, teams built the same integration 16 times for proprietary APIs.

From Reference Servers to Enterprise Firewalls

▶ Watch (2:27)

MCP launched in November 2024 with reference servers for Postgres, SQLite, Git, and file systems. SaaS companies followed: ClickHouse, Notion, and Zapier built integrations. Creative uses emerged for Blender, Ableton, 3D printers, and a Fantasy Premier League server. Behind corporate firewalls, teams wired MCP to Salesforce, Jira, Snowflake, and HR systems. Bishop said these internal deployments represent the largest but least visible MCP adoption surface.

The 2026 Productionization Roadmap

▶ Watch (10:38)

Bishop called 2025 the year MCP proved it was needed. 2026 is about productionizing agentic systems. Three protocol changes land in the June specification. A new stateless transport lets hyperscalers scale MCP like standard web services. The tasks primitive supports long-running agentic work. Cross-app access ties into corporate identity providers so users skip OAuth flows. Bishop credited Google and Microsoft for helping shape the transport redesign.

Better Clients Fix Context Bloat

▶ Watch (17:57)

Bishop identified client quality as the root cause of most MCP criticism. Context bloat happens when clients dump all tools into the window. The fix is progressive discovery: load tools only when needed through tool search. Claude Code reduced tool overhead from 22% of a 200K token window to near zero. Bishop also urged composability through code, giving the model an interpreter to call MCP via structured outputs as an alternative to CLI-style piping.

Notable Quotes

Behind every corporate firewall, we’re quietly wiring MCPs to systems of records and to the company data all day long. Diamond Bishop · ▶ Watch (4:28)

2025 was about figuring out if a thing like MCP um is needed in the ecosystem. I think the answer is clearly the resounding yes. Diamond Bishop · ▶ Watch (9:40)

We need to go and build better clients. If there’s one thing that I’m really worried about and really, you know, maybe complaining about um in the ecosystem, it’s the current state of clients. Diamond Bishop · ▶ Watch (17:57)

The best part is that, hopefully, we’re still humble enough to say, “Okay, we got it wrong. Let’s go and fix it together.” Diamond Bishop · ▶ Watch (22:08)

Key Takeaways

  • MCP reached 110 million monthly SDK downloads in 16 months, outpacing React’s trajectory.
  • Enterprise adoption behind firewalls represents MCP’s largest but least visible deployment surface.
  • The June 2026 specification adds stateless transport, long-running tasks, and cross-app access.
  • Progressive discovery eliminates context bloat by loading tools only when needed.

About the Speaker(s)

Diamond Bishop leads Datadog’s AI Skunkworks group. He joined Datadog through the acquisition of Augmend, the DevOps AI assistant startup he co-founded. Diamond has over 15 years of AI experience building assistants and developer platforms, from the early days of Microsoft Cortana and Amazon Alexa to open-source work on PyTorch at Meta. Today he focuses on unhobbling enterprise agents: durable control flow, strong eval and…