First, the thing no other page will tell you
No one publishes a price for this, not a single reputable company.
Instead, you will find many similar articles, mostly from development agencies, all with titles like "MCP server development cost." Their numbers vary widely and do not match up.
| Published by | Price range |
|---|---|
| One agency | $3,000 to $60,000 |
| One advisory firm | $100,000 to more than $1,000,000 |
When published prices for the same deliverable differ by more than three hundred times, they are not talking about the same work. Averaging them would be meaningless, so this page will not do that. Instead, it explains what really affects the cost, which is more helpful.
The one thing that decides the price
Do you already have a good API?
If you do, adding the agent layer is a small job. The transport part is already solved with an official SDK, and you can get a first version working in a day. The real work is choosing which ten or twenty operations to turn into tools, writing clear descriptions, keeping results within size limits, and building a way to check if a model uses them correctly. This takes weeks, not months, and most of the work is design, not coding.
If you do not, this is not really an agent project. It is an API project that ends with an agent endpoint. You need stable identifiers, proper authentication, clear error handling, and rate limiting before you can add a protocol. This is a much bigger job, and the agent part is just a small piece of it.
So, the honest answer to the cost depends on something you can figure out in an afternoon: could a skilled stranger use your API just by reading the documentation?
What agent-ready actually means
I have published a standard for this called Wired for Agents, and it is intentionally brief. There are four criteria: pass or fail, with no scores or maturity levels. You can read it for free and decide if you agree.
- A documented API. Public, versioned, with a machine-readable OpenAPI description at a stable URL that needs no login to read.
- An MCP server whose tools meaningfully overlap that API, each with a valid input schema.
- Programmatic authentication. Scoped, revocable machine credentials. No human in the loop at the moment of the call.
- Machine discovery. An agent starting from your bare domain can find the doors without logging in, and every URL you advertise actually resolves.
Most products miss the fourth requirement, but it is also the easiest and cheapest to fix. Discovery usually just needs a few files and clear information about which URLs actually work.
Where the time actually goes
There are four main tasks, and none of them is about the protocol itself.
- You need to decide what not to expose. Creating a tool for every endpoint is a common mistake. Having eighty tools is worse than none, because the descriptions overwhelm the model and it makes more mistakes. Careful selection is key.
- You also need to protect write actions. Anything that sends, pays, deletes, or publishes should assume the instruction might come from a document, not a person. This means adding confirmation steps and limiting permissions.
- Evaluation is also important. You need to run real tasks and check if the model picked the right tool, not just if the endpoint returned a 200 status. Most cost estimates skip this step, but it is what proves your setup works.
- Staying up to date is also part of the work. The specification was revised on 28 July 2026 and removed sessions, the handshake, and stream resumability, and deprecated three features. Anything built before then needs updates. There is now a twelve-month deprecation window, which helps, but things still change often.
Three cheaper things to try first
Here they are in order from cheapest to most expensive. Two of them do not require any custom code.
- First, fix discovery. If your documentation is behind a login or your published URLs return 404 errors, no protocol can fix that. This is often just an afternoon's work.
- Next, enable it at the gateway. If your API is already behind Azure API Management, Google Cloud API Gateway, or Amazon Bedrock AgentCore, these services can expose your OpenAPI-described API as an MCP server without custom code. You still need to choose which operations to expose, but you avoid building from scratch.
- Another option is to generate it from the specification. SDK generators can create an MCP server along with your client libraries. The main providers no longer publish prices, so you will need to discuss costs with them, but this is a real option.
When to do none of this
For most products, agent traffic is still not significant. It is growing fast, and visitors from AI referrals tend to convert better than average, but the total numbers are still small and no reliable source shares the full data.
Be cautious with the statistic that says more than half of web traffic is automated. While true, it refers to crawlers like scrapers and search indexers. It does not mean agents are calling your product for customers, so using this to sell you an MCP server is misleading.
So, if your API is not solid yet, focus on improving it. That investment pays off even if agents never use it, which is not true for the server. If your product is just public content, a clear API and honest documentation give you most of the benefits. And if your team cannot support evaluation and maintenance, an unchecked server will have problems you may never notice.
Anthropic, which co-founded the protocol, says this in its own build guide: if you already have a public API or CLI that Claude can use directly, you can leave the connector out. When the party with the most to gain from adoption publishes the case for skipping it, that is worth more than my opinion.
What I charge, since I am asking you to publish yours
I price this the same way as everything else here. A Diagnostic Week costs $2,500 and includes a written scope, a plan, a fixed quote for the build, and a clear report of any remaining issues. Builds usually cost between $15,000 and $50,000, based on the Diagnostic Week, not a guess.
I do not list a fixed price for an MCP server, because the real cost depends on your API. Quoting before reviewing the code often leads to estimates that are far too low.