Will MCP Become HTTP for Agents—or Another Plugin Protocol?

Calling MCP the HTTP of AI captures its ambition: clients discover resources and invoke tools through a common contract, while service providers avoid bespoke integrations for every agent. The skeptical view is that it becomes another plugin protocol fragmented by platform extensions. Both outcomes remain possible.

The MCP specification defines clients, servers, tools, resources, prompts, transports, authorization, and security guidance. Its immediate value is practical: it reduces an N-by-M integration problem into relatively stable implementations on each side.

HTTP succeeded through more than message syntax. Addressing, caching, proxies, identity, browsers, and servers formed an ecosystem. A common interface does not create common trust. MCP still needs robust discovery, compatibility, permissions, signed distribution, and observability.

Tool providers should not expose existing APIs without redesign. Names, schemas, errors, and idempotency must work for agents; consequential tools need previews and explicit side effects. Machine-callable does not mean safely machine-understandable.

Clients face a connector supply-chain problem. A server can solicit excessive authorization, return prompt-injection content, or overstate capabilities. The MCP security boundary belongs in the host, not the connector. Sandboxing, least privilege, provenance, and runtime logs are host responsibilities.

Standardization will commoditize basic connectors and move value toward domain semantics, reliable operations, and organizational policy. Teaching an agent how to complete an expense process correctly remains more valuable than merely exposing a create-record endpoint.

Enterprises should begin with an owned allowlist: server identity, owner, permissions, data destination, and version policy. Calls need attribution to a specific user and task rather than a shared omnipotent credential. Open ecosystems and governance are complementary.

MCP becomes foundational only if models, clients, and services can be replaced at low cost. Its best future is infrastructure for substitutability. Stable host boundaries would make it more than a plugin system; proprietary divergence would recreate the adapter work it set out to remove.