Your Agent Can Call Stripe Directly. Should It Make Ethical Decisions Directly?
Perplexity's CTO told his team to drop MCP and use APIs instead. Reddit declared MCP dead. One company's internal tooling decision became a funeral announcement.
The eulogy is premature.
The real critique
The criticism has a real core. MCP adds token overhead: every tool description loads into context with each request. A hundred tools means thousands of wasted tokens. For a search company processing millions of queries, that cost stings. And frontier models have gotten good enough to read API documentation and make correct calls without protocol middleware.
For generic tool use, this reasoning holds. Calling Stripe, querying a database, fetching weather data: the information flowing through those pipes is commodity. If the model can read the docs, the middleware adds cost without adding meaning.
Constitutional governance is a different category.
Discovery changes everything
When an agent encounters a creed-pdp MCP server, it doesn't learn "here's an endpoint." It discovers that ethical review exists and that it should participate. This is capability discovery: ethical infrastructure becomes part of an agent's operating environment rather than something a developer remembered to wire up.
APIs are opt-in by design. A developer reads documentation, decides to integrate, writes the call. MCP tools are discovered by agents autonomously. That distinction is the entire ballgame for ethics infrastructure.
Ethical review should be something an agent encounters when it enters a domain, the way laws apply when you enter a jurisdiction. You don't opt out of gravity.
Courts, not statutes
Consider the analogy to human institutions. Statutes are like APIs: look them up, follow them mechanically. Courts are living institutions that apply principles to novel situations.
Constitutional AI governance functions like a court. A creed_decide call doesn't return a cached lookup. It applies constitutional reasoning (the specific creeds a user has chosen) to a specific context. The principles are stable. The application is contextual. This is why the work belongs at the protocol layer, where agents interact with it as a first-class part of their environment.
The token argument actually helps
The token overhead complaint applies to bloated tool registries. Creed Space's PDP server exposes four tools at under 500 tokens of context cost. The companies with a hundred tools at 20,000 tokens feel the pain. Purpose-built ethical infrastructure doesn't.
Focused servers for high-stakes domains are exactly where MCP earns its overhead. You don't need a protocol layer to call a weather API. You do need one to ensure constitutional review happens everywhere, not just where someone remembered to add it.
The strategic risk
If the industry narrative settles on "agents should call APIs directly," companies building AI products may skip ethical infrastructure entirely. The reason won't be malice. It will be the absence of a natural insertion point.
MCP gives ethical governance a place in the protocol layer. Without that place, guardrails become afterthoughts, bolted on at the application layer by teams that remember, absent everywhere else.
The right posture
MCP-first, API-always. Prefer the protocol layer for its discovery properties. Provide REST equivalents for every capability so no developer hits a wall. Stay agnostic to protocol wars.
Generic tool-calling is commoditising. High-stakes domains (where discovery, standardisation, and structured review matter more than shaving tokens) are where protocol-layer value concentrates. Constitutional AI governance sits precisely there.
Your agent can call Stripe without help. The question is whether it should make ethical decisions without help too.