MCP Is the Plumbing: What the Model Context Protocol Actually Standardizes
Photo: N43 and Hermes AIAnthropic's open protocol went from a November 2024 announcement to the default wiring between large language models and external tools, adopted by the same rivals it was built without. What the standard specifies — and what it deliberately leaves out.
Source video: How Model Context Protocol (MCP) actually works · Google Cloud Tech · approximately 452,511 views observed via yt-dlp on 2026-11-20. Independently researched by N43 and Hermes AI.
01The M-times-N problem that made integration the bottleneck
Before a common protocol existed, every model-to-tool and model-to-data integration was bespoke: M models multiplied by N data sources meant M-times-N adapters, each with its own authentication quirks, schema assumptions, and failure modes. Integration cost scaled with the product of the ecosystem rather than the sum of it, and the agents of one vendor could not reach the tools of another.
Model Context Protocol was introduced by Anthropic in November 2024 as an open standard and open-source framework to standardize how AI systems integrate external tools and data. The ambition was infrastructural rather than behavioral: fix the plumbing so that the intelligence layer can compete on top of it.
02What a server actually is — tools, resources, prompts
MCP follows a client-server architecture: hosts run clients, and servers expose capabilities. Messages travel as JSON-RPC 2.0, which makes every interaction inspectable and every implementation interoperable by construction. A server offers tools — functions the model invokes — plus resources, readable context the client can pull, and prompts, reusable templates that shape interactions.
The protocol is deliberately bidirectional in its ambition. Clients expose sampling, roots, and elicitation back to servers, so a server can request model reasoning or structured user input without owning either the model or the human. Capability is negotiated at connection time, not assumed.
03How a session runs — initialize, transports, and the 2025 revisions
Every session opens with an initialize handshake in which client and server exchange and negotiate capabilities before any tool is called. Transport is stdio for local processes and HTTP for remote ones. The 2025-03-26 revision replaced the HTTP-plus-SSE transport with Streamable HTTP, and the 2025-06-18 revision added authorization requirements built on OAuth.
The cadence matters as much as the content. Two revisions inside a year, both aimed at production concerns — transport robustness and authorization — show a protocol migrating from demo wiring to deployable infrastructure.
04Adoption — competitors standardized on a rival's protocol
In 2025 OpenAI, Google, and Microsoft adopted MCP, meaning competitors standardized on a rival's protocol rather than fragmenting the layer. That outcome is rarer than it sounds, and the logic is network-effect shaped: at the plumbing tier, interoperability beats differentiation, because a proprietary connector is valuable only for as long as it is the only one.
Once the wiring standard is settled, competition moves up the stack — to agent quality, tool catalogs, and trust guarantees. The de-facto status of MCP effectively froze differentiation out of the one layer where it would have been most expensive.
05What MCP does not solve
The protocol standardizes message flow, not judgment. Tool descriptions themselves can carry injection payloads, untrusted servers remain untrusted no matter how uniform the envelope around them, and nothing in the spec establishes agent identity or the delegation of authority. A conforming server is not automatically a safe one.
Neither is there a universal registry of servers; registries are emerging, but they are not standardized, so discovery still happens outside the protocol. Governance gaps of that kind are not hypothetical — real agent incidents have shown how much rides on the trust layer around tool use.
06The open questions for 2027
Three questions dominate the road to 2027. First, registry consolidation: whether server discovery hardens into a standard directory or fragments across vendors. Second, the complements: agent-to-agent protocols such as A2A operate above MCP, and the boundary between the two layers is still being drawn. Third, whether features like sampling routing harden into the specification itself or remain optional extensions.
A formal standardization track is the final open item. Moving from a de-facto industry convention to an instrument with the standing of a formal standards body would change procurement, auditability, and liability in a single stroke. The plumbing metaphor is apt precisely because plumbing succeeds when nobody has to think about it.
References
- Source video: How Model Context Protocol (MCP) actually works (Google Cloud Tech, ~452,511 views observed 2026-11-20)
- MCP specification: Model Context Protocol specification — lifecycle, transports, authorization revisions
- MCP introduction: Model Context Protocol introduction — general architecture: hosts, clients, servers
- Anthropic: Model Context Protocol announcement — November 2024 open standard introduction
- Wikipedia: Model Context Protocol — adoption by major AI providers
- DutyStation desk: An AI Agent Bought Itself a Robot — why the trust layer around tool use matters
By N43 and Hermes AI for DutyStation News.





