← Carlos Cardona

When you actually need an MCP

Every few weeks I end up in the same debate: are MCPs ready for production, should we build one, are they hype. I think the debate stays confused because it skips the only question that matters: what does an MCP actually add over a REST API or a CLI?

The uncomfortable answer is: no capability at all. Anything an agent can do through an MCP server, it could do through an API call or a shell command. What MCP adds is standardized self-description. A compliant server tells any compliant client, at runtime, what tools exist, what their schemas are, and how to call them. The integration work on the client side drops to zero. That is the entire value proposition, and it is a real one, but it is a distribution feature, not a capability.

That framing gives you a clean decision test: who consumes this?

If you control both the agent and the service, you do not need an MCP. Call your API directly with tool definitions, or let the agent use your CLI. If, on the other hand, arbitrary AI clients you do not control should be able to plug into your service, Claude, IDEs, other people's agents, then MCP is exactly the right tool, because the alternative is shipping and maintaining an integration per client.

And the CLI option deserves more respect than it gets, because a visible trend is forming around it. Modern agents live in a real shell. Once an agent can run gh or wrangler directly, wrapping the same functionality in an MCP server is a detour: extra schemas, extra round trips, extra tokens. Several teams are landing on a hybrid that I also use in my own Cloudflare setup: keep the capability in the CLI, and use a thin layer whose only job is to teach the agent how to use it. Anthropic pushed this furthest with code execution with MCP, where the agent writes code against servers instead of loading every tool definition into context; their worked example dropped from roughly 150,000 tokens to about 2,000. When the agent can already act, description beats wrapping.

So why not MCP for everything? Because of the part nobody solved yet: identity.

A RESTful API comes with decades of settled machinery. Users, sessions, scopes, middleware that knows exactly who is asking and what they are allowed to touch. Now ask the same question in MCP world: when an agent connects on my behalf, who does the server think is calling, and what stops it from seeing everything? This was the strongest objection in that meetup debate, and it holds. The authorization spec leans on OAuth extensions that identity providers implement inconsistently, to the point that some enterprise voices call it a non-starter as it stands. In practice, deployments accumulate long-lived tokens with no revocation path, which is precisely the attack vector you do not want multiplying. The emerging best practice, scoping access per tool rather than per server, is sensible and still young.

So my current rules are these. If a well-known CLI already does the job, let the agent use it. If you control both sides, call your API and skip the ceremony. Build an MCP server when you are exposing a capability to AI clients you do not control, keep it a thin, well-described wrapper over a normal API where your real auth lives, and treat anything touching per-user permissions as the risk surface it is. MCPs are not hype and they are not the new everything. They are a standard for the boundary, and boundaries are worth standardizing.