MCP: the standard that connects AI to your information system

Connecting an AI assistant to a CRM, an ERP or a document base still follows the same pattern in most organizations: one custom integration project per "AI tool × system" pair. Anthropic put it plainly when it published MCP on November 25, 2024: every new data source requires its own implementation. With N systems and M AI tools, the bill grows as N×M.

MCP replaces those N×M integrations with N+M standard connectors. Each system exposes an MCP server — the socket. Each AI tool ships an MCP client — the plug. The official documentation claims the USB-C analogy outright: one standard port replaces the proprietary cables. The arithmetic checks out on a typical case.

The difference changes the slope of the cost curve: every new AI tool meant re-plumbing everything you already had; now it costs one connector.

A standard is worth what its adopters make of it — starting with its creator's competitors. The facts are dated and public.

The foundation counts eight platinum members: AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI — head-on competitors funding the same infrastructure, as with USB or HTTP. Usage follows: more than 10,000 published servers by December 2025, SDK downloads growing from 97 million per month to nearly half a billion between December 2025 and July 2026. Honeycomb, an observability vendor, reported in July 2026 that close to 20% of its monthly interactive queries came from agents. Betting on this standard is now less risky than multiplying proprietary integrations.

MCP standardizes the connection, not the trust. The specification dedicates an entire document to protocol-specific attacks: the "confused deputy" problem on intermediary servers, an explicit ban on forwarding access tokens issued for other services, session hijacking. An academic study published with ACM catalogs 16 threat scenarios across a server's lifecycle. A standard connector is not a safe connector.

Second blind spot: the official registry, still in preview as of summer 2026, does not list private servers. The internal catalog — which system exposes what, to whom, with which rights — remains the IT department's job, with a recommendation to host your own private registry.

Third point: the protocol is still moving. Three major revisions between November 2025 and July 2026, including a shift to a "stateless" architecture that changes how servers are built. Deprecation is now guaranteed for at least twelve months — real predictability, which still requires budgeting connector maintenance. This is the kind of trade-off we document in how we build.

What the status quo costs — Every proprietary integration launched today is an asset you will rewrite tomorrow: 50 projects where 15 connectors would do, a gap that widens with every system you add — and the connectors you fund as custom work often already exist among the 10,000 servers published by late 2025.

These seven questions structure the decision; they are part of the scoping work we run in our services.

The N+M reasoning assumes a minimum of volume: with a single AI tool and two systems, the immediate gain is small. The protocol evolves fast, and integration savings are not maintenance savings. Finally, MCP says nothing about data quality, business-level permissions or the relevance of AI answers: cleanly plugging into a poorly governed IT estate mostly produces errors faster. The connection is standard; the value remains specific to each company.