$19/mo for an MCP server is the right instinct — but monthly pricing is the wrong primitive for agents

The Whoff Agents team is right that $19/month lands in the "instant decision" zone for a working developer. The psychology is correct: below $25, most engineers don't wait for budget approval. They put it on a card and expense it later.

The problem is that agents aren't engineers. And monthly subscriptions aren't the right pricing model for workloads that vary by 100x.

The usage distribution problem

A human developer paying $19/month for an MCP server integration uses it consistently. They build a workflow, it runs daily, the value is clear.

An AI agent using the same server has a completely different usage pattern. A research agent might call your server 200 times in one hour while working on a project, then go dark for three days. A customer support agent might call once per conversation, but run 10,000 conversations a month.

Monthly subscription pricing fails for both of those cases. The research agent gets an incredible deal (200 calls for $19). The customer support team gets price-shocked when they realize $19/month for their volume would need to be $19/agent/conversation to make financial sense.

The right primitive for MCP servers isn't monthly recurring — it's per-call. The price should scale with usage, not time.

Why per-call billing is hard to build

The reason most MCP servers default to free or monthly tiers isn't lack of ambition. It's that per-call billing requires infrastructure that doesn't exist in the MCP protocol itself:

• How does the calling agent know the price before it calls?
• How does the server authorize and reserve payment before running?
• How does settlement confirm without a round trip that adds latency to every call?
• What happens when an agent runs out of credit mid-workflow?

Whoff's instinct to charge is the right move. The execution model for how to charge is what the ecosystem is still figuring out.

The Agent FICO layer

MnemoPay introduces an Agent FICO score (300-850) — a creditworthiness signal on the calling agent side that determines what tools it can access and at what price. On the server side, the SDK exposes pricing per tool, handles authorization before execution, and confirms settlement in the return payload.

The net effect: developers building MCP servers can set per-call prices, the payment infrastructure runs at protocol level rather than as a bolt-on, and agents understand their own credit standing before making expensive calls.

672 tests in v1.0.0-beta.1, npm-native, listed on Smithery and ClawHub. if you're building MCP infrastructure and pricing is part of the architecture conversation, the per-call model is worth a look: https://mnemopay.com