How to Monetize an MCP Server: Charging AI Agents for Tool Calls
Key takeaways
- There are three common ways to charge for an MCP server: a subscription, a price per call, and a price per purchase or unit of work.
- The MCP specification defines authorization through OAuth but defines no payment step, so a paid tool needs a payment layer outside the protocol.
- Very small per-call charges are hard on card rails, because each card transaction carries fixed costs and every charge can be disputed.
- An agent should not spend a user's money without the user's approval, so a paid tool needs an approval step that names the amount.
- A common pattern is that a paid tool answers with a price, the user approves the named amount, the charge runs, and only then does the tool run.
To monetize an MCP server, put a price on the tools that do valuable work, and make sure a human approves each charge before the tool runs. The Model Context Protocol (MCP) lets an AI agent discover and call tools on a server, and its specification covers authorization through OAuth. It does not define how a tool asks for payment, so a server owner adds that step with a payment layer of their choice.
What are the ways to charge for an MCP server?
The three common ways to charge for an MCP server are a subscription, a price per call, and a price per purchase. The right choice depends on who your caller is and how much value a single call delivers.
| Model | How it works | Fits best when | Main problem |
|---|---|---|---|
| Subscription | The user pays a recurring fee for access, and calls are unlimited or capped. | The same user calls your server often and predictably. | An agent that needs one call must first sign a person up for a plan. |
| Per call | Each tool call costs a small fixed amount. | Each call is cheap to serve and usage varies widely. | Tiny charges are expensive and slow to settle on card rails. |
| Per purchase | A tool call buys a specific outcome, such as a booking, an order or a report, at a stated price. | One call delivers clear value, such as a booking or a finished document. | The user must approve each amount, which needs an approval step. |
Why is per-call billing hard on card rails?
Per-call billing is hard on card rails because card payments were designed for purchases, not for fractions of a cent. Three properties of card rails make very small per-call charges difficult.
- Each transaction has fixed costs: A card charge carries costs that do not shrink with the amount. When the price of a call is tiny, those costs can be larger than the price itself.
- Each charge can be disputed: A user can dispute any card charge through their bank. Thousands of tiny charges create thousands of possible disputes, each with its own handling cost.
- Users do not read thousands of lines: A statement full of small charges from one server is hard for a person to check, and charges that a person cannot recognize get disputed.
Two practical answers exist today. The first is to price units of work that are large enough to stand as a card purchase, and to let one approval cover several calls, so the user sees one total and one charge. The second is a non-card rail. The x402 protocol, created by Coinbase and now hosted by the Linux Foundation, uses the HTTP 402 Payment Required status code so that a server can ask for a payment, usually in stablecoins, before it answers a request.
Why must users approve what an agent spends?
Users must approve what an agent spends because the agent acts for a person, and the person owns the card. If the agent can spend without approval, every mistake the agent makes becomes a charge the person did not want, and that charge becomes a dispute against the server owner. An approval that names the amount protects all three parties.
- The user: Sees what the agent wants to buy and at what price, and can refuse.
- The agent: Cannot approve on the user's behalf, so it cannot spend more than the person agreed to.
- The server owner: Gets a record of what the user approved, when, and what was charged, which can be used to answer a dispute.
Card network programs for agents, such as Visa Intelligent Commerce and Mastercard Agent Pay, follow the same principle: the cardholder authenticates, often with a passkey, and the agent transacts within the limits the cardholder set.
How does an MCP server get paid by agents?
An MCP server gets paid by agents when a paid tool returns a price, the user approves that amount, a payment layer charges the user, and the tool then runs. The steps below describe the general pattern.
- 1
Price your tools: Give a paid tool a fixed price, or a quote that your code calculates when the agent calls. Only your code sets the price.
- 2
The agent calls the tool: Instead of running, the tool answers with an offer: the work and the price. Nothing is charged yet.
- 3
The user approves: The agent shows the offer to the user, who approves the named amount on a surface the agent cannot control, such as a wallet or an approval page with a passkey.
- 4
The charge runs: The payment layer charges the user's card or other payment method. Tokenization keeps the card number away from the agent and from your server.
- 5
Your tool runs: Your tool runs with proof of the approved purchase, and your handler does the work. A fair design charges the user only for work that completed.
You can keep your tool code and your framework. You decide which tools are free, which are open to guests, which need a sign-in, and which cost money. Before live payments, test the full flow in a sandbox that charges no real money.
Should a paid tool use a fixed price or a quote?
A paid tool should use a fixed price when every call does the same work, and a quote when the price depends on the request.
- Use a fixed price: When every call does the same work, such as one report or one lookup. The agent and the user see the same price every time.
- Use a quote: When the price depends on the request, such as the size of a document or the dates of a booking. Your code calculates the amount, and the quote expires after a short time.
- Refuse when you cannot sell: A quote can decline a request it cannot serve, so the user is never asked to approve work you will not do.
- Keep quotes free of side effects: A quote may be asked for more than once before approval, so it must calculate a price and nothing else.
How do agents find a paid MCP server?
Agents find MCP servers through registries and search indexes that list public servers and their tools. The MCP project maintains an official registry, and several third-party indexes exist. Clear tool names, accurate descriptions and a stated price help an agent pick the right tool for a task.
This page is a general overview. It is not legal or tax advice. Dispute outcomes depend on card network rules, issuer policy and the agreements between the parties.
Agents are a new kind of customer for an MCP server, and they can pay when the person behind them approves. Price the work your tools do, let the user approve each amount, and let the charge run on rails that users and banks already trust.
Frequently asked questions
How do I charge for an MCP server?
To charge for an MCP server, choose a pricing model (subscription, per call, or per purchase), put a price on the tools that do valuable work, and add a payment layer that gets the user's approval of each amount before the paid tool runs.
Does the MCP specification include payments?
No. The MCP specification defines how agents discover and call tools and how servers authorize clients through OAuth, but it defines no payment step. A server owner adds payments with a separate payment layer or protocol.
Can I charge per tool call?
Yes, you can charge per tool call, but very small card charges are costly because each card transaction carries fixed costs and each charge can be disputed. Common answers are to let one approval cover several priced calls, or to use a non-card rail such as the x402 protocol.
Why does the user have to approve each purchase?
The user has to approve each purchase because the user owns the payment method and the agent only acts for them. An approval that names the amount stops an agent from spending more than the person agreed to, and gives the server owner a record to answer a dispute.
Do I need to rewrite my MCP server to accept payments?
No. A server owner can usually keep the existing tool code and framework, add a payment step in front of the paid tools, and set a price or a quote on each one. Free tools stay free.
Does an MCP server need to see the user's card number?
No. With tokenization, an MCP server receives proof of an approved purchase and never the card number, which keeps the server out of most card data handling.
Related articles