Six ways to earn from MCP servers, with starting price ranges, three worked examples with the math, running costs, common margin mistakes, and a 30 day plan to land a first paying customer, from per-call billing to monthly retainers and custom builds.
Most people building MCP servers give them away. They publish a repo, collect a few stars, and the only number that moves is the hosting bill for their test server. A smaller group is charging real money for the same kind of software, and the difference is rarely the code. It comes down to the pricing model, the buyer they picked, and whether the server saves someone an hour they were already paying for.
This article lays out where MCP server income comes from, six pricing models with their trade-offs, three worked examples with the math, what running a server costs, and a 30 day plan to land a first paying customer. The price ranges are starting points to test, not market data, because the market is young and nobody publishes a reliable price list yet.
💡 Quick definition: MCP (Model Context Protocol) is an open protocol, introduced by Anthropic in November 2024, that lets AI assistants call external tools and read external data through one standard interface. An MCP server is the program that exposes those tools.
Nobody wakes up wanting an MCP server. A bakery owner wants orders to stop getting lost between chat messages and a spreadsheet. An analyst wants last quarter's filings summarized before a Monday meeting. A support lead wants the assistant to read the refund policy and the order history in the same breath. The server is plumbing. What you sell is the pipe that finally connects an assistant to the system where the work lives.
That changes how you describe the product. "A Postgres MCP server with six tools" sells nothing. "Ask which invoices are 30 days late and get the answer in ten seconds" sells a lot. Write your offer in the buyer's words, then work backward to the tool list.
Three groups pay most reliably:
Consumers rarely pay. Businesses do, because the hour they save has a dollar value attached.
Before you quote anything, put a number on the buyer's pain. Say a six person support team spends three hours a week each copying order details between browser tabs. At $40 an hour, that is $2,880 a month of paid time spent on copy and paste. A $2,400 build and a $250 retainer look cheap beside that figure, and you can say so in one sentence on the first call.
Ask three questions in every first call: how many people do this task, how often, and what does an hour of their time cost? Multiply the answers, then price your server at a small fraction of the result. Ten to twenty percent of the monthly time saved is a reasonable ceiling to test, because the buyer keeps most of the benefit and the deal feels like an easy yes.
Each model suits a different buyer and a different support load. Here is the side-by-side before the details.
| Model | Starting range to test | Best for | Main risk |
|---|---|---|---|
| Pay per call | $0.005 to $0.05 per call | Developer-facing hosted servers | Revenue swings with usage |
| Monthly subscription | $9 to $99 per month | Small teams with steady use | Churn when novelty fades |
| Flat-fee custom build | $1,500 to $15,000 once | Businesses with one clear workflow | Scope creep |
| Monthly retainer | $300 to $2,500 per month | Clients who rely on the server daily | Support hours creeping up |
| Marketplace listing | Free to $20 per month | Visibility and inbound leads | Thin direct revenue |
| Free tier plus upgrade | $0, then $19 to $149 | Wide reach and fast feedback | Free users cost real money |
Meter every tool call and bill for what gets used. This works when your server wraps something that costs money each time it runs, such as a licensed data source, an image generation job, or a heavy scraping task. It needs a remote server where each customer has their own access token, so you can count calls per customer.
Price at three to five times your cost per call. If a call costs you $0.004 in compute and upstream fees, charging $0.02 leaves room for failed calls, free trials, and your own time. Buyers dislike surprise invoices, so offer a monthly spending cap they can set themselves.
A flat monthly fee with an included quota is the easiest model to explain. Three tiers is plenty: a small one for individuals, a middle one for teams, and a large one that makes the middle one look sensible. Include a call allowance in each, and charge a modest overage rate instead of cutting people off mid-task.
Subscriptions give you predictable revenue and give buyers a predictable budget line. The risk is churn once the novelty fades, so tie the plan to a recurring job, like a weekly report, not an occasional curiosity. Offering two free months on a yearly plan improves your cash flow and keeps people from reconsidering every thirty days.
You build one server for one company around one workflow, and quote a fixed price after a paid scoping call. The deliverable includes the server, deployment, short documentation, and a handover call. It is the fastest route to first income because the buyer already has the problem and a budget.
Write the tool list into the contract. Every extra tool becomes a priced change request, which is the cheapest protection against scope creep you will ever buy.
A retainer pairs naturally with a custom build. It pays for hosting, monitoring, fixes when an upstream API changes, and new tools as the client's needs grow. Tokens expire, vendors rename fields, and somebody will message you at 7 am saying the assistant stopped finding orders. The retainer pays for that message.
One workable starting point is 10 to 20 percent of the build fee per month. Track your support hours for a quarter, then adjust.
Public directories of MCP servers exist, and listing a server there costs little or nothing. Treat a listing as advertising, not income. Point it at a paid hosted version, a team plan, or a contact form for custom work. Direct revenue from listings alone is thin, but a good listing brings inbound leads while you sleep.
Open source the server, then charge for hosting, team features, and support. Developers try it for free, and the ones who do not want to run it themselves pay. The danger is that hosted free users cost real money, so cap the free tier at a number of monthly calls that costs you less than a cup of coffee.
💡 Transport matters for billing. A server that runs locally over stdio lives on the buyer's machine, so you cannot meter it. Per-call billing and subscriptions need a remote server over HTTP with authentication. For local servers, sell licenses, support, or custom work instead.
These are illustrative scenarios built to show the arithmetic. They are not reported results from real businesses.
A freelancer builds a custom server for a neighborhood bakery. It has three tools: list today's orders, mark an order ready, and check flour stock in a spreadsheet. The build takes 25 hours. The quote is $2,400 flat, plus a $250 monthly retainer for hosting and fixes.
| Line | Amount |
|---|---|
| Build fee | $2,400 |
| Retainer, 12 months | $3,000 |
| Year one revenue | $5,400 |
| Hosting, 12 months | $180 |
| Time: build plus 2 hours a month of support | 49 hours |
| Effective rate after hosting | about $106 per hour |
The retainer makes up more than half of year one revenue. That is why a one-off job becomes a small business only when you attach recurring work to it.
A founder licenses a filings dataset and wraps it in a hosted server with three tools: search filings, pull a table, and compare two periods. Three subscription tiers: $29 for 500 calls, $79 for 3,000 calls, and $199 for 12,000 calls. With 40 customers split 20, 15, and 5 across the tiers:
The data license is the moat. Anyone can write three tools in a weekend. Far fewer people can get the data legally, and buyers pay for access they cannot easily reproduce.
This one bills per call. A marketing agency wants its assistant to turn a product description into finished images. The server takes the description, calls an image API, waits for the result, and returns a hosted URL. Suppose the upstream image cost is $0.02 per image (check your provider's current price) and you charge $0.08. A client making 3,000 images a month pays $240. Upstream cost is $60 and hosting is $20, which leaves about $160.
PicassoIA's developer API suits this job. It works like Replicate: you create a job, poll it, then fetch the result. It exposes four models, including PicassoIA Image and PicassoIA Image Editor Pro. Each account can run 5 predictions at once, so a busy server needs a queue, and a promise like "100 images in a minute" needs testing before it goes into a contract. Plan terms for API access can change, so check the current details on the PicassoIA site before you commit to margins.
A small server fits a $5 to $20 monthly container or a serverless function. Watch cold starts: an assistant that waits eight seconds for a tool call looks broken, so keep one instance warm. Add a free uptime check and an alert that reaches your phone. Budget $20 to $60 a month for your first few customers.
Authentication deserves the same care as uptime. Issue each customer their own credential so you can revoke one without breaking the rest, and log every call with a timestamp, the tool name, and a status code. Those logs are your invoice backup, your debugging trail, and your proof when a client insists the assistant never used the tool.
Your server may call a model itself to summarize or classify, or call a paid third-party API. Decide early whether you pass those costs through or bake them into the price. A fast, inexpensive model such as Gemini 3.5 Flash suits routine steps like sorting a support ticket, while a stronger model such as Claude Sonnet 5 is better for multi-step reasoning. Test both on 50 real inputs before you pick.
| Cost line | Monthly estimate | Note |
|---|---|---|
| Hosting | $5 to $60 | One warm instance per region |
| Monitoring | $0 to $15 | Free checks are enough at first |
| Upstream APIs | Varies | Pass through or bake into price |
| Payment processing | About 3 percent plus a small fixed fee | Most card processors work this way |
| Support time | 1 to 4 hours per client | The biggest hidden cost |
💡 Price your own hours. Most new MCP builders forget support time. If a $250 retainer eats four hours a month, you earn about $60 an hour before hosting. Raise the price or cut the support scope.
An assistant decides which tool to call by reading its description. Weak descriptions cause wrong calls, and wrong calls cause refund requests. Claude Sonnet 5 on PicassoIA is a quick way to draft and tighten those descriptions, plus your README, pricing page copy, and onboarding emails.
Two parameter tips matter. Give the exact parameter names so the model does not invent fields, and keep descriptions short, because clients send every description to the model on each request. Long descriptions cost tokens on every single call.
| Week | Goal | Output |
|---|---|---|
| 1 | Pick one buyer and one painful workflow | Five short conversations, one written problem statement |
| 2 | Build 3 to 5 tools | A working server on a remote host with authentication |
| 3 | Pilot with one buyer | A discounted trial, call logs, a list of fixes |
| 4 | Set prices and ask for the sale | A contract, an invoice, and a retainer offer |
Keep the first offer narrow. A server that does one job well, for one type of buyer, is easier to price, easier to support, and easier to describe in a single sentence. Once two or three clients run on it, you can package it as a subscription and stop starting from zero each time.
💡 Aim for recurring first. A $2,400 build feels good for a week. A $250 retainer that runs for two years is worth $6,000. Whenever you quote a build, attach a monthly line.
Every MCP business needs words and visuals: a pricing page, a launch post, screenshots, a banner image for the README. Picasso IA puts the models for all of it in one place. Draft your tool descriptions with Claude Sonnet 5, then create photorealistic visuals for your landing page with Seedream 5 Pro.
Pick one pricing model from the table above, write the offer in your buyer's words, and test it this week. Then open Picasso IA, try a model on your own idea, and see how fast a rough plan turns into something you can show a client.
