Pricing AI Agents and MCP Servers: Models and Examples

To price an AI agent or MCP server, start with where it runs and what each use costs you. If it runs on the buyer's machine and costs you nothing per use, a one-time price or a simple subscription for updates works well. If it runs on your servers or calls a paid model, usage-based pricing, credits or a subscription with usage limits protects you from losing money on heavy users. If your agent produces a clear, measurable result, such as a resolved support ticket, you can also charge per outcome.

After that, pick a price based on the value to the buyer, check it against similar products, and test it. There is no correct number to look up. This guide covers the main pricing models, real published examples, how to pick between them, how anchoring and free tiers work, and a simple process for setting your first price. It is part of our wider guide on how to sell AI agents, MCP servers and skills.

Start with your costs

Write down what each user costs you before you think about what they will pay.

• Fixed costs: your time to build and maintain, domain, docs hosting.
• Per-user costs: hosting, database storage, support time.
• Per-use costs: model API calls, third-party API calls, compute per job.

A local MCP server or a skill pack has almost no per-use cost. A hosted agent that calls a large model on every request may cost you real money each time it runs. That difference matters more than anything else in this article. If per-use costs are high and you charge a flat fee, your most active users can cost more than they pay. Our comparison of local vs remote MCP servers explains why that split exists.

What are the main ways to price an AI agent?

Model	Buyer pays	Best for	Watch out for
One-time	Once, up front	Downloads, skill packs, local servers, templates	No recurring income; you still fix bugs
Subscription	Monthly or yearly	Hosted tools, products with frequent updates, team features	Buyers expect ongoing value and support
Usage-based	Per call, per job or per result	Hosted agents with variable compute or model costs	Unpredictable bills worry buyers
Credits	Prepaid credits spent on different actions	Products with several kinds of usage at different costs	Buyers must learn what a credit buys
Outcome-based	Per resolved task or other agreed result	Agents with a clear, measurable result	Disputes over what counts as an outcome

One-time pricing

The buyer pays once and keeps what they bought. This is the easiest model to explain and the easiest to buy. It fits anything that runs locally and changes slowly: a skill pack, a configured agent template, a bundled MCP server.

To keep income going, sell major versions separately or offer a period of free updates, such as twelve months, after which a renewal is optional. Be clear about what the price includes so buyers are not surprised later.

Subscription pricing

The buyer pays monthly or yearly. Subscriptions match products that keep delivering value: a hosted MCP server, an agent that runs scheduled jobs, or a tool with regular new features. Yearly plans are often offered at a discount compared to paying monthly, which gives you cash up front and the buyer a lower total.

Subscriptions raise expectations. Buyers who pay every month expect uptime, fixes and answers. Only choose this model if you can keep shipping.

Usage-based pricing

The buyer pays for what they use: per tool call, per task, per document processed. This keeps your margin steady when costs scale with use. Some marketplaces support it directly. Apify Store's pay-per-event model, for example, lets developers define billable events and set a price for each, and the developer receives 80% of that revenue. Per-call payments over HTTP are also possible with x402, which uses the HTTP 402 Payment Required status code.

The risk is buyer anxiety. People dislike open-ended bills, especially for agents that can loop or retry. Reduce that by showing clear unit prices, letting users set spending caps, and sending alerts before a cap is reached.

Credit-based pricing

The buyer prepays for a bundle of credits, and each action spends some of them. Credits let you price different actions at different rates, for example one credit for a search and ten for a long report, without a separate price list for each. Ibbaka's analysis of AI agent pricing describes credits that stand for tokens, time or actions, sold in fixed bundles or pay-as-you-go, with choices about whether unused credits expire or roll over. Make the conversion rate visible, or buyers cannot tell what they are paying for.

Outcome-based pricing

The buyer pays only when the agent achieves an agreed result, such as a resolved support conversation. Intercom prices its Fin agent this way, at $0.99 per outcome, and counts an outcome when a customer confirms the issue is resolved, does not ask for more help, or Fin completes a workflow. Ibbaka argues outcome pricing works best when the outcome is clear and agreed, the agent's contribution can be attributed, and outcomes are predictable. Most agents will not meet all three, so outcome pricing is often combined with a usage fee.

Not every vendor prefers it. Decagon offers per-conversation and per-resolution pricing and says most of its customers pick per-conversation because it is predictable and avoids arguments over what counts as a resolution.

Hybrid pricing

Many hosted products combine models: a monthly fee that includes a set amount of usage, plus a per-unit price above it. This gives buyers a predictable base and protects you from heavy users. It is a sensible default for hosted agents that call paid models.

How do the models compare?

This table is a qualitative summary of the trade-offs above, not measured data.

Model	Risk for the buyer	Revenue predictability for you	Margin protection when costs rise
One-time	Pays up front, before seeing ongoing value	Low; sales arrive in bursts	Poor if each use costs you money
Subscription	Pays whether or not they use it	High	Poor for heavy users unless usage is capped
Usage-based	Bills can grow if an agent loops	Medium; follows usage	Good
Credits	Prepays and may lose unused credits	Medium to high	Good, if each action's credit cost tracks your cost
Outcome-based	Low; pays only for results	Low; depends on success rate	Mixed; failed attempts cost you and earn nothing
Hybrid (base fee plus usage)	Medium	High base, variable top-up	Good

Real pricing examples

Published prices, checked on 2026-11-12. Check each source before you copy a number, because prices change.

Product or platform	Model	Published price or split	Source
Intercom Fin	Outcome-based	$0.99 per outcome	Intercom pricing
Decagon	Per conversation or per resolution	No public prices; per-resolution is a higher fixed rate	Decagon on pricing AI agents
Apify Store Actors	Pay per event or pay per usage	Apify keeps 20% of Actor revenue	Apify monetization docs
Gumroad (selling a download)	One-time or subscription	10% + $0.50 per direct sale; 30% via Discover	Gumroad pricing

How to choose a price

There is no formula, but a short process helps.

1. Estimate the value. What does your tool save the buyer? Time, a paid service, a manual process? A developer tool that saves an hour a week is worth more than one that saves a click.
2. Look at alternatives. Check what similar agents, servers and SaaS tools charge, and what doing it by hand would cost. Your price should make sense next to those.
3. Cover your costs with room to spare. For usage-heavy products, make sure your price per unit is well above your cost per unit, including payment and platform fees.
4. Pick a round, simple number. Simple prices are easier to understand and compare.
5. Write down your reasoning. You will revise the price later, and you need to know what you assumed.

Remember that marketplace and payment fees come off the top. AgenticMart's paid selling is not open yet; the planned platform fee is 12%. Other platforms and payment providers have their own fees; the guide to selling MCP servers and agents compares them. Factor them in before you decide your price.

Anchoring

Anchoring is the effect where the first number a buyer sees shapes how they judge the next one. In practice, this means showing a higher-priced option next to your main one.

Ways to use it honestly:

• Offer a team or pro tier above your main tier. Some buyers will pick it, and it makes the main tier look reasonable.
• Show the yearly price as a monthly equivalent next to the monthly price.
• Compare your price to the real cost of the alternative, such as the hours saved.

Do not invent fake original prices or discounts. Misleading price claims can break consumer protection rules.

Should you have a free tier?

For agents and MCP servers, a free option is often the strongest trust builder you have. Buyers want to see the code work, and check that it is safe, before they pay.

Common free options:

• Free open-source core. The base project is free and open; pro features are paid. See going from open source to a paid agent for how to split features fairly.
• Free usage allowance. A hosted server includes a small monthly allowance.
• Free trial. Full access for a limited time.
• Free lite version. A smaller skill pack or a server with fewer tools.

A free tier must be useful on its own, or it creates frustration instead of trust. It also creates support load, so set limits you can sustain. Make the upgrade path obvious: show what paid adds, in plain words, at the moment a user hits a limit.

Pricing and trust go together

Buyers weigh price against risk. Publishing a manifest, declaring network domains and sharing a security report lowers that perceived risk. A free public scan report is one way to do that. On AgenticMart, any listing can get a free automated scan; each listing page shows whether it has one (see how the scan works). A scan is a malware and risk check, not a safety guarantee. AgenticMart's developer tools, an API, an MCP server and a CLI, show how listings and their scan status appear to other tools. The AI agent security guide covers what security-minded buyers look for.

Test and adjust

Your first price is a guess. Treat it that way.

• Ask early buyers why they bought and what almost stopped them.
• Grandfather existing customers when you raise prices. It keeps goodwill and costs little.
• Watch support load per customer. A low price that brings heavy support may not be worth it.

FAQ

What is the best pricing model for an MCP server?

For a local MCP server, a one-time price or a free open-source core with a paid pro bundle is usually simplest. For a hosted MCP server, a subscription with included usage works well. If each call costs you real money, add per-unit pricing above the included amount.

Should I charge per tool call?

Charge per call only if your costs scale with calls and buyers can predict their use. Otherwise, buyers may worry about runaway bills from agents that loop. If you do charge per call, offer spending caps and alerts.

What is outcome-based pricing for AI agents?

You charge only when the agent delivers an agreed result, such as a resolved support conversation. Intercom's Fin, at $0.99 per outcome, is a published example. It works when the outcome is easy to measure and attribute to the agent.

Are credits better than plain usage pricing?

Credits help when your product has several actions with different costs, because one unit covers all of them. Plain per-call pricing is easier to understand when there is only one kind of action.

How do I raise prices without upsetting users?

Give notice, explain what has improved, and keep existing customers on their current price for a while or for good. Raising prices for new buyers only is the gentlest option. Being clear and early matters more than the size of the change.