How to Sell MCP Servers, AI Agents and Agent Skills

To sell an MCP server, AI agent or agent skill, pick where the code will run, pick one business model (paid download, open core, hosted service, support and licensing, or sponsorships), then choose a sales route whose fee and tax handling you can live with. Code that runs on the buyer's machine is easy to copy, so it usually sells as a download, a pro bundle or a license. Code that runs on your servers sells as a subscription or per use.

This guide is for developers who build the code. If you run an agency and want to sell agent services to businesses, the advice on contracts and SLAs below still applies, but the business models are different. Below you will find each model, a comparison of sales routes and their fees as published today, licensing in brief, the trust signals buyers check, and the basics of payments and tax.

What are you actually selling?

Before you choose a model, be precise about the product. The three common types behave differently when someone pays for them.

• MCP servers expose tools, resources and prompts to an AI client through the Model Context Protocol. They can run locally (the client starts them as a process) or remotely (the client connects over HTTP). If you are new to the format, see what an MCP server is.
• Agent skills are folders with a SKILL.md file of instructions, plus optional scripts and reference files. They are mostly text and small scripts, so they are easy to copy. Our page on what agent skills are covers the basics.
• Agents are complete programs that plan and act, often combining a model, tools and memory. They range from a CLI tool to a full hosted product.

The key question is where the code runs. Code that runs on the buyer's machine is easy to copy once delivered. Code that runs on your servers is not, but you pay for hosting and you carry uptime and data-handling duties. For the trade-offs, read local vs remote MCP servers.

Which business model fits an MCP server or agent?

Model	How you get paid	Works best when	Main downside
Paid download	One-time payment for a zip or package	Self-contained tools, templates, skill packs	Easy to share once bought; no recurring income
Open core	Free base, paid pro features or add-ons	You already have open-source users	Deciding what is free vs paid causes friction
Hosted (remote) service	Subscription or usage-based billing	Your server holds the value (data, compute, integrations)	You run infrastructure and handle user data
Support and licensing	Support contracts or commercial licenses	Companies depend on your code	Sales cycles are slower and more personal
Sponsorships	Recurring donations or sponsor tiers	Widely used infrastructure projects	Income is voluntary and hard to predict

Paid download

You sell a packaged deliverable: a skill pack, an MCP server bundle, an agent template with prompts and config. The buyer pays once and downloads a file. You do not run servers and you do not track usage. Anything delivered can be copied, so buyers are paying for convenience, curation, documentation and updates. Consider charging for major new versions.

Open core

You keep a useful base project open source and sell extras: advanced tools, team features, integrations with paid APIs, or a polished bundle with setup help. The hard part is drawing the line between free and paid without making existing users feel the project is being taken away. We cover that in the open core model for AI agents.

Hosted or remote service

You run the agent or MCP server yourself and users connect to it. The MCP specification supports remote servers over HTTP with authorization, so a paid remote server is a normal product shape. You can charge a subscription, charge by usage, or both. You also take on hosting, model API bills if your server calls an LLM, monitoring, and a duty to protect the data users send you.

Support and commercial licensing

If companies already run your open-source code, some will pay for guaranteed response times, private help, or a license with different terms. A common version is dual licensing: the public code is under a copyleft license like AGPL-3.0, and companies that do not want those obligations buy a commercial license. You can only do this if you hold the rights to all the code, which usually means contributors sign a contributor license agreement.

Sponsorships

GitHub Sponsors and Open Collective let users and companies fund your work on a recurring basis. Sponsorships suit infrastructure that many people rely on but that is hard to sell directly. Offer clear tiers, such as a logo in the README or a monthly office hour, so sponsors know what they are paying for.

Models stack well: an open-source server can add sponsorships, then a paid pro bundle, then a hosted version. Start with the model that fits the product you have today.

Where can you sell an MCP server? Routes and fees compared

You have three broad options: your own site, a general digital-goods platform, or an agent and MCP marketplace or directory. Most sellers use more than one. The table lists fees as each provider publishes them, checked on 2026-11-12. Fees change, so open the source before you decide.

Route	Fee to you (published)	Who handles sales tax	Sells MCP servers or agents?	Source
Apify Store	Apify keeps 20% of Actor revenue	Check Apify's terms	Yes, Actors; its pay-per-event docs list MCP compatibility	Apify monetization docs
Gumroad (direct link)	10% + $0.50 per transaction	Gumroad, as merchant of record	Any digital file	Gumroad pricing
Gumroad (Discover marketplace)	30% per transaction	Gumroad, as merchant of record	Any digital file	Gumroad pricing
Lemon Squeezy	5% + $0.50 per transaction, some payments extra	Lemon Squeezy, as merchant of record	Any digital product or subscription	Lemon Squeezy pricing
GitHub Sponsors	0% from personal accounts; up to 6% from organisation accounts	Not a sale of goods	Funds the project, not a product sale	About GitHub Sponsors
Official MCP Registry	Free; lists metadata only, no payments	Not applicable	Discovery only	MCP Registry
AgenticMart	Free to list. Paid selling is not open yet; the planned platform fee is 12%	To be confirmed before checkout opens	Listing and discovery today; paid sales planned	Selling on AgenticMart
Your own site	Your payment processor's fee only	You, unless you use a merchant of record	Anything you can deliver	Your processor's pricing page

Some notes on reading the table:

• Your own site gives you full control over pricing, branding and the customer relationship. You need a payment provider, a way to deliver files or issue API keys, and your own traffic. Hosted products usually end up here, because you need account management anyway.
• A merchant of record is the legal seller and handles sales tax or VAT collection. Gumroad and Lemon Squeezy say on their pricing pages that they act as one. That convenience is part of what their fee pays for.
• General platforms are not agent-specific. Buyers will not find you there by browsing for MCP servers, so you still need discovery.
• Agent and MCP directories bring buyers who are already looking. On AgenticMart you can list free in the MCP server directory or the agent skills directory, get a storefront at /@handle, and any listing can get a free automated scan; each listing page shows whether it has one. If someone already imported your public repo, you can claim the listing by signing in with GitHub, as long as you have admin or maintain permission on the repo. Getting listed in the right places is its own skill, covered in how to get your MCP server discovered.

No marketplace guarantees sales. Treat each listing as one channel among several.

Which license should you sell under?

Your license decides what buyers and users may legally do with your code. MIT and Apache-2.0 allow anyone to use and resell your code, AGPL-3.0 requires network users to be offered modified source, and paid downloads usually ship under a custom commercial license. The full comparison, including dual licensing and the relicensing history of open core companies, is in the open core model for AI agents. This is not legal advice; talk to a software licensing lawyer before you sell commercial licenses.

How should you price it?

Match the price model to your costs: one-time prices suit downloads, subscriptions suit products that keep running or changing, and per-use or outcome pricing suits hosted agents whose costs scale with use. Subtract the route's fee from the table above before you set a number. The full breakdown, with real examples, is in pricing AI agents and MCP servers.

Contracts, SLAs and liability

Once you sell to companies, especially a hosted server, buyers will ask about terms as well as price. This is general information, not legal advice.

• Terms of service or a license agreement. State what the buyer gets, how many seats or installs, whether they may redistribute, and how updates work.
• A service level agreement (SLA) for hosted products. Only promise uptime and response times you can measure and meet. An SLA usually names the remedy, such as service credits, if you miss it.
• Limitation of liability. Agents act on the buyer's files, accounts and APIs. Your terms should say what you are and are not responsible for, and you should get advice on how far a cap can go where you sell.
• Data processing. If your hosted server receives customer data, buyers may ask where it is stored, who can access it and how long you keep it. Write this down before the first enterprise buyer asks.
• Security questions. Expect a security questionnaire from larger buyers. A public scan report and a permissions manifest answer many of its questions in advance.

How do buyers decide whether to trust you?

Buying an agent or MCP server is not like buying a font. The buyer is handing your code access to their files, their tokens and their network. Many developers have read about tool poisoning in MCP and supply-chain attacks on AI agents, and they check before they install.

You can lower that barrier with a few concrete steps:

1. Declare what you need. List the environment variables, file paths and network domains your code uses. A server.json, SKILL.md or agent manifest makes this machine-readable.
2. Ask for the least access. Do not request write access when read is enough, or a broad token when a scoped one works. Our guide to least-privilege AI agents explains why buyers look for this.
3. Keep secrets out of the repo. Run a secret scanner before every release.
4. Pin and audit dependencies. Known vulnerabilities in dependencies are an easy reason for a buyer to say no.
5. Make builds reproducible. If a buyer can rebuild what you shipped, they can check that the download matches the source.
6. Publish a security report. A third-party, public report lets buyers check your claims instead of taking your word for it.

On AgenticMart, a scan runs on a fresh virtual machine. It combines Semgrep static analysis, Gitleaks secret detection and OSV-Scanner dependency checks, tries to install and start the code, records outbound connections and compares them to the domains you declared. The result is a public report for one exact commit. A scan is a malware and risk check, not a safety guarantee. You can read how the scan works. For the buyer's side, see the AI agent security checklist and the AI agent security guide.

Payments and taxes: the basics

This section is general information, not tax or legal advice. Rules differ by country and change over time.

Payments and payouts

Marketplaces that let you sell usually pay you through a connected account at a payment processor. With Stripe Connect Express accounts, for example, Stripe handles onboarding and identity verification, and you see balances and payouts in a Stripe dashboard. Expect to give identity and bank details to the processor, not to the marketplace. Availability depends on your country, so check before you plan around one provider.

Some agents charge per call over HTTP instead of through a checkout, using x402, a protocol built on the HTTP 402 Payment Required status code. That fits a hosted MCP server billed by use.

Sales tax and VAT

Many places tax digital products and software services, and some tax them based on where the buyer lives. A merchant-of-record platform is the legal seller and generally handles it. When you sell directly, or through a marketplace that is not the merchant of record, the responsibility may fall on you. Tools such as Stripe Tax can help calculate tax, but you still need to know your obligations.

Income tax and records

Sales income is usually taxable. Keep records of each sale, each fee and each payout from day one. If you sell regularly, you may need to register a business. An accountant can tell you what applies to you.

A practical launch checklist

1. Decide where your code runs and pick one business model to start.
2. Choose a license and add it to the repo.
3. Write a README that explains what it does, who it is for and how to install it in under five minutes.
4. Add a manifest that declares permissions, env vars and network domains.
5. Run secret, dependency and static analysis scans, and fix what they find.
6. Pick a sales route from the table and note its fee and tax handling.
7. Set a price, and write down why you chose it.
8. List in the registries and directories that fit your product. AgenticMart's developer tools include an API, an MCP server and a CLI for checking how your listing appears.
9. Write terms, and an SLA if you host the product.
10. Ship updates on a visible schedule and keep a changelog.

FAQ

Can I sell an MCP server that is open source?

Yes. Open-source licenses like MIT and Apache-2.0 allow selling. Many developers keep the code free and charge for a hosted version, a pro bundle, support or a commercial license. Buyers are usually paying for convenience, updates and help rather than for the code alone.

Where can I sell an MCP server today?

You can sell from your own site, through a general digital-goods platform such as Gumroad or Lemon Squeezy, or as an Actor on Apify Store. The official MCP Registry and most directories, including AgenticMart for now, list servers but do not take payments. AgenticMart's paid selling is planned, with a planned 12% platform fee.

What fees do MCP and agent marketplaces charge?

As published on 2026-11-12: Apify keeps 20% of Actor revenue, Gumroad charges 10% plus $0.50 on direct sales and 30% on Discover sales, Lemon Squeezy charges 5% plus $0.50, and GitHub Sponsors takes nothing from personal sponsorships and up to 6% from organisations. Check each provider's page before you decide, because fees change.

Which is better for agents, a one-time price or a subscription?

It depends on your costs and how often the product changes. If your code runs on the buyer's machine and changes rarely, a one-time price is simple and fair. If you host it, pay for model calls, or ship frequent updates, a subscription or usage-based price fits better.

Do I need a company to sell AI agents?

Not always, but it depends on your country and how much you sell. Many payment providers let individuals sign up, but tax and registration rules vary. Check your local rules or ask an accountant before you start selling regularly.

How do I convince buyers my agent is safe to run?

Declare the access your code needs, keep that access minimal, remove secrets from the repo and keep dependencies patched. Then make it checkable: publish a manifest and a public scan report so buyers can verify your claims instead of trusting them. A scan lowers risk but does not prove code is safe.

Do I need an SLA to sell a hosted MCP server?

Not by law in most places, but business buyers often ask for one. Only promise uptime and response times you can measure and meet, and state the remedy if you miss them.