How much does a custom MCP server cost? The drivers that decide the price

Why there is no list price

Searching for the price of a custom MCP server turns up numbers that are hard to compare, because the phrase covers very different jobs. A server that lets one person read one clean API from their laptop and a server that lets a whole company read and update three systems through Claude and ChatGPT are both custom MCP servers. The protocol is the same. The work is not.

The useful question is not what a server costs but what your server needs to do. Once you can answer the five questions below, any builder, in-house or external, can give you a real estimate, and you can compare quotes on the same scope.

Driver 1: how many systems, and how clean their APIs are

Every source system adds its own authentication, its own rate limits, its own data model, and its own failure modes. One system with a modern, documented API is the cheapest case. Several systems that describe the same customer or order differently cost more, because the server has to decide which source wins when they disagree, or show both values with their sources.

Systems with no API at all, such as an old database or a shared spreadsheet, need a governed read path or a sync into a store the server can query. That is often the single largest line in a quote.

• Count the systems, not the features you want.
• Ask which ones have a usable API and who owns access to it.
• Note any place where two systems hold the same data with different values.

Driver 2: the auth model

The cheapest setup is one shared credential, such as an API key that every request uses. Anthropic's documentation is clear about the trade-off: a connector that uses a shared credential gives everyone who uses it that credential's access. For low-risk data that may be acceptable. For anything with customer, financial, or HR data, it usually is not.

Per-user access means OAuth. The MCP authorization specification is built on OAuth 2.1, and requires the server to validate that each access token was issued for it. Doing this properly, connecting it to your identity provider, and checking each user's permissions in the source system on every call is real engineering work. It is also the work that gets a connector through a security review.

Driver 3: read-only or write actions

A read-only server answers questions. A server with write actions creates tickets, updates records, or sends messages. Writes cost more to build because each one needs strict input validation, clear labeling so the assistant asks before using it, protection against a retried request running twice, and tests for the cases where it must refuse.

Both Anthropic and OpenAI let tools be marked by behavior. OpenAI's plugin documentation describes readOnlyHint, destructiveHint, and openWorldHint annotations, and says they help ChatGPT choose confirmation behavior but do not replace authorization, validation, or confirmation in your server. Budget for that server-side work on every write tool.

Driver 4: hosting and network path

Claude reaches custom connectors from Anthropic's cloud, not from the user's device, so the server must be reachable from the public internet or from Anthropic's published IP ranges through your firewall. ChatGPT needs a public HTTPS endpoint or OpenAI's Secure MCP Tunnel. Either way you are running a production web service with certificates, uptime monitoring, logging, and someone on call.

Hosting in a vendor's environment is usually cheaper to start. Hosting inside your own cloud account, often required by security teams for sensitive data, adds deployment and handover work.

Driver 5: maintenance after launch

The first version is rarely the expensive part over a year. Source system APIs change, Claude and ChatGPT change how they handle tools and admin settings, the MCP specification is revised, and people ask for new tools once the first ones work. Ask any builder what is included after launch and how changes are priced.

For usage-priced bridges, model your volume. Zapier's documentation says Zapier MCP runs on your existing Zapier plan with no separate MCP billing, and each successful tool call uses two tasks from your plan's allowance. A question that searches and then updates ten records is eleven tool calls. At low volume that is cheap; at team scale, compare it with a fixed-scope build.

How to get a quote you can compare

Two quotes for the same server can differ widely because they assume different things. Before you compare numbers, check what each one includes, so you are not choosing a cheaper build that leaves the expensive parts to your team.

Write down the ten or twenty questions people will ask, the systems that hold the answers, whether each person should see only their own data, which actions should be allowed, and where the server must run. Send the same list to every option you are weighing. Prestyj quotes custom MCP servers this way: a fixed written scope after a short scoping call, with hosting and maintenance priced separately so you can see both.

• Does the price include OAuth with per-user permissions, or a shared key?
• Which AI clients will it be tested in: Claude, ChatGPT, or both?
• Who hosts it, and who is alerted when it goes down at night?
• How many write tools are included, and how is each one confirmed and tested?
• What happens when a source system changes its API or the MCP specification is revised?
• Do you own the code and configuration, and can you move it in-house later?
• How are new tools priced after launch?