How MCP servers are priced, and what you are actually paying for
Free, subscription, one-time licence — and the metered model that keeps being tried and keeps disappearing. What each one does to a buyer, and what it does to a publisher.
Pricing here is not like pricing SaaS. The marginal cost of a call is real but tiny. The value is large and occasional. And the buyer is often not a person at all but an agent, which is a customer with no judgement and a great deal of enthusiasm.
Given that, you would expect a mess of experiments. What the market has actually converged on is three models and a graveyard.
Why per-call billing keeps failing
Metering only works when whoever bills you also executes the call and owns the meter. When the server runs on its publisher’s hardware and something else proxies it, the invoice is a number two parties can disagree about — and one month, someone’s retry loop will make them disagree loudly.
An allowance says the same thing without the argument: a number of calls a month, and past it the call is refused rather than repriced. Both sides know the figure in advance and neither has to trust the other’s counter.
A limit that keeps serving past the limit is not a limit. It is a surprise with a delay on it.
Which leaves the question every plan has to answer eventually: what happens at the edge of it. There are only two answers, and they produce very different customers.
Which brings us to the part publishers argue with hardest, and the part the evidence is clearest on: what a free tier is actually for.
What each model is really for
- Free is a distribution decision, not a discount. Its job is to remove the evaluation, because the evaluation is the expensive part for the buyer and the fatal part for the publisher.
- A subscription is the natural shape for a remote server: there is an admission point, so entitlement can be checked, and an allowance can be enforced.
- A one-time licence is the natural shape for a local one, for exactly the opposite reason: there is nothing to check, so a recurring charge has nothing behind it.
How to read a price before you install
| What you see | What to check |
|---|---|
| “Free tier” | How many calls, and what happens at the end of them |
| A monthly price | The allowance attached to it, and whether it resets or rolls |
| “Unlimited” | The rate limit, which is where the real ceiling is hiding |
| A one-time licence | Whether updates are included, and for how long |
| Per-call pricing | Whose counter decides, and what you can see of it |
If you are pricing one
- Open free, with a real allowance rather than a trial. A trial is a deadline; an allowance is a habit.
- Sell the next tier on the allowance, not on features. Splitting features across tiers means the model has to know which tier it is on, and it does not.
- Mind the floor. Card processing has a flat component, so pricing at a couple of dollars a month hands a meaningful share of every sale to the processor before anybody’s percentage.
- Then leave the price alone. Once a free tier exists, price is the lever that moves installs least — and the one publishers reach for first.
The uncomfortable version, for publishers: the install is not the product. The second month is. Most of what looks like a pricing problem is a server that never got wired into anything, and no price fixes that.