The short answer

A custom MCP server that connects AI assistants to one live business system, with authentication and permission scoping, typically costs $15k–$40k and takes 3–4 weeks for a read-only first version. Servers with write operations, approval workflows and several connected systems take 6–8 weeks and cost more, because the systems behind the server, not the protocol, drive the work.

MCP (Model Context Protocol) turned "connect our AI to our systems" from a custom integration per assistant into one server that every compatible client can use. That makes the server worth building properly, and it makes the question "what does it cost?" worth answering honestly. These are the ranges we quote, and why.

What drives MCP server cost

The protocol layer is the cheap part: open-source SDKs handle the message format. Almost all of the cost sits in five decisions:

Typical ranges by scope

ScopeTypical rangeTimeline
Connector into one live business system, with auth and permission scoping $15k–$40k 3–4 weeks for a read-only first version
Several systems, write operations and approval workflows Above the single-system range; quoted after discovery 6–8 weeks
ERPNext AI rollout: MCP server plus chat assistant and several automated document types (see ERPNext AI integration) $30k–$80k Depends on document types
Claude setup: a handful of skills, one or two connectors and a team plugin $25k–$60k Skills in 1–2 weeks; connectors 3–6 weeks

Ongoing cost is small by comparison: hosting for the server itself, plus maintenance when the connected systems change their APIs. Model usage is paid by whichever AI client calls the server, not by the server.

Auth, permissions and audit: the part cheap quotes leave out

A demo MCP server can be built in an afternoon. The difference between that and one you can point at production is security design, and it is most of what you are paying for:

If a quote does not mention these, ask where they are. They are much cheaper to design in than to add after an incident.

Hosting options

Most business MCP servers run in the client's own cloud account, next to the systems they connect to, which keeps data inside existing security boundaries. On-premise deployment behind a firewall is common for ERPs that are not internet-facing. Managed hosting suits teams without an operations function. The choice affects setup work more than ongoing cost.

Timeline

A read-only server over one or two systems typically reaches production in 3–4 weeks: discovery and tool design in the first week, build and permission work in the next two, testing against real queries in the last. Adding write operations, approval workflows and more systems takes it to 6–8 weeks. The variable is almost always the systems being connected, not the protocol.

How to keep the cost down

For how we build and scope these, see our custom MCP server development services. If you are new to the protocol, start with what MCP is in AI.

Common questions

How much does a custom MCP server cost?

A custom MCP server connecting AI assistants to one live business system, including authentication and permission scoping, typically costs $15,000–$40,000. Servers with write operations, approval workflows and several systems cost more and take 6–8 weeks instead of 3–4.

How long does it take to build an MCP server?

A read-only server over one or two systems takes about 3–4 weeks. Servers with write operations, approval workflows and several systems take 6–8 weeks. The connected systems drive the timeline, not the protocol.

Why do MCP server quotes vary so much?

Because the protocol is the cheap part. Quotes differ mainly on how many systems are connected, whether the AI may write as well as read, and how much security design is included: per-user permissions, approval gates and audit logging.

Is it cheaper to build one MCP server or separate integrations for each AI tool?

One MCP server, as soon as more than one AI client needs the same systems. Separate integrations repeat the build and the permission logic for every assistant; an MCP server builds both once and every compatible client can use it.

Related reading