The Model Context Protocol started as a simple way for AI agents to reach tools. Today, the MCP Dev Summit opens at the University of Toronto with 70 speakers, six tracks, and fifty-plus sessions — the first major in-person gathering of the protocol layer builders who turned an experiment into the foundation of the agent economy. The official schedule makes the scale visible, but the more important question is not how many people showed up. It is whether this community will address the coordination gap between the layers it has built.
The numbers are no longer small. MCP’s Tier 1 SDKs now pull approximately 500 million downloads per month, up from 97 million in March, with over one billion lifetime downloads across Python and TypeScript. The protocol’s public server count sits near 15,930 across four major registries. Google’s A2A protocol, now under the Linux Foundation with over 150 organizations in production, addresses horizontal agent-to-agent communication — the direction MCP explicitly does not cover. AP2 added payments. Okta’s XAA brought agent identity to the provider layer with 25-plus early adopters. Each protocol solved its piece. None of them coordinate.
That coordination gap is where the summit’s agenda gets interesting. The TELUS keynote — Kevin Ng and Dana Harrison presenting “One Year of MCP in the Enterprise: Registry, Workflows, Agents, and Remote Hosting” — is the clearest production-deployment voice on the schedule. TELUS has run MCP in production for a full year across enterprise workflows, and their presence signals that the protocol has survived enough operational reality to teach the room something about what actually connects and what does not. Michael Forrester’s Accenture keynote, “Governing MCP for a Workforce the Size of a City,” carries a similar weight: 200,000 deployed users make the coordination problem concrete at enterprise scale.
The tensions live in the workshop tracks. Johns Hopkins University is running “A2A Meets MCP: Tool Integration vs.” — the “versus” in the title doing its own work, because A2A and MCP solve adjacent problems and no shared coordination fabric connects them. Google’s Daryl Ducharme is running “One Router for Your Agent’s MCP Servers: Aggregate, Authorize and Observe Every Tool Call” on Day Two, which is essentially a workaround for the missing coordination layer: build a single router that manages what the protocols do not manage together. CrowdStrike’s Daniel Davidson is presenting on insider threats in the agent stack — the security surface expanding because each protocol layer introduces its own trust boundary and no meta-layer enforces policy across them.
This is the same 80-point gap our Agent Orchestration Gap analysis documented: Cisco RSA data showed 85 percent of enterprises experimenting with AI agents but only 5 percent in production. The bottleneck is not model capability or infrastructure availability. It is the absence of a coordination layer that ties MCP tool access, A2A horizontal routing, K8s sandbox execution, execution-layer gateways, XAA identity, and IETF identity drafts into a single enforceable fabric. Each layer works. The integration between them does not.
The MCP Dev Summit has the right people in the room. The question is whether convening four protocol communities in one conference center produces proposals for the connective tissue — or whether the summit validates each layer in isolation and leaves the coordination gap to whoever can build around it next.
Sources (verified this cycle):
Named in prose without linking (schedule confirmed, URLs not independently verifiable): TELUS keynote (Kevin Ng + Dana Harrison), Accenture keynote (Michael Forrester), Johns Hopkins workshop (Andrew Nicklin + Maeve Mulholland), Google workshop (Daryl Ducharme), CrowdStrike session (Daniel Davidson).
