How much does it cost to build a custom MCP server?

At Reflekt Lab, scoping one workflow costs $2,300 to $4,700, and a first MCP server in production, covering one workflow with its tools, permissions and per-user sign-in, starts at $17,500. Hosting is about $50 to $100 a month, and running it, with monitoring and updates when the specification or your upstream APIs change, starts at about $470 a month. Agency cost guides published in 2026 range from about $3,000 for a simple single-tool server to $90,000 and more for several systems with write access.

How long does it take to build an MCP server?

Typically 2 to 4 weeks from a signed scope to a first workflow in production. A local server that wraps one read-only API can run in a day with the official SDKs; the weeks go into sign-in, permissions, safe writes, hosting and testing with real users.

What does an MCP server cost to run?

The hosting is small: about $50 to $100 a month for the servers we run for clients. The server has no AI model bill of its own, because the model runs in your users' Claude or ChatGPT. What costs money is keeping it working: the MCP specification was revised three times in 2025 and again in July 2026, and your own APIs change too. Our running plans start at about $470 a month.

What makes an MCP server more expensive to build?

Four things: how many systems it connects to, whether each user signs in as themselves (OAuth) or the server uses one shared key, whether the AI can change records or only read them, and whether the system has a clean API at all. A read-only server over one modern API is the cheap end; writes to a legacy ERP with per-user permissions are the expensive end.

Do I need a custom MCP server if an official one exists?

Often not. Salesforce, QuickBooks Online, GitHub and many SaaS tools now ship official MCP servers or connectors. Use the official one when it covers your workflow. A custom server is worth it when the official one is missing the actions you need, when your system has no server at all (an in-house tool, a legacy ERP, a database), or when you need several systems behind one set of permissions.

Can I build an MCP server myself?

Yes, and for an internal, read-only tool it is often the right call. The official SDKs make a local server a short project; our step-by-step guide builds one in Go. Production is the harder part: remote hosting, OAuth sign-in, permissions per user, previews before writes, and monitoring.

MCP Server Development Cost: Our Real Prices (2026)

An isometric server block cabled to a database, a filing cabinet and a stack of documents, next to a stack of coins and a gauge

Short answer: at Reflekt Lab, a custom MCP server costs $2,300 to $4,700 to scope, from $17,500 to put a first workflow into production, and from about $470 a month to run, plus about $50 to $100 a month of hosting. A first workflow typically takes 2 to 4 weeks from a signed scope. The server itself has no AI model bill: the model runs in your users' Claude or ChatGPT. What makes one expensive is not the protocol. It is the number of systems behind it, how users sign in, whether the AI may change records, and how clean the systems' APIs are. Below: our prices, what the market quotes, what moves the number, the running cost, and when you should not build one at all.

Our prices

These are the same prices as on our MCP page, in US dollars.

Step	What you get	Price
Scoping (one workflow)	A half-day on what the AI should be able to do, which systems and actions are in, who may do what, and a fixed price and date for the build	$2,300 to $4,700
A first workflow in production	The MCP server with its tools, per-user sign-in, permissions, previews before every write, hosting set up, tested with your users	from $17,500
Hosting	The server on our platform, HTTPS, logs	about $50 to $100 a month
Running it	Monitoring, updates when the MCP specification or your APIs change, a monthly allocation of changes	from about $470 a month

Delivery is typically 2 to 4 weeks from a signed scope to a first workflow in production. Each further workflow on the same server costs less and goes faster, because the sign-in, permissions and hosting are already in place. We run around 195 tools on one server of our own, for our CRM, an order bot and internal agents (six of the servers we run).

What the market quotes

The 2026 cost guides published by development agencies spread wide: from about $3,000 to $8,000 for a simple single-tool server, through $8,000 to $25,000 and more once it has authentication and permissions, to $35,000 to $90,000 for several systems with write access, and six figures for enterprise and regulated builds (Bacancy, AppMatic, Satekk). One guide budgets maintenance at 15% to 30% of the build each year.

Read those numbers with care, ours included: every one comes from a firm that sells the work, and "an MCP server" covers anything from one read-only tool on a laptop to a multi-tenant service with OAuth. Before comparing quotes, compare the scope: how many tools, how users sign in, whether it writes, and who runs it afterwards.

What moves the price

Four things, roughly in this order:

How many systems sit behind it. One modern API is cheap. A CRM, an ERP and a spreadsheet behind one server means three integrations, each with its own authentication, limits and failure modes.
How users sign in. A server that uses one shared key is simple, and unsafe the moment several people use it. A remote server where each person signs in as themselves, with their own permissions, needs a proper OAuth flow (what MCP OAuth involves). For anything more than one person, it is not optional.
Whether the AI can change things. Reading is cheap to make safe. Writing needs previews before every change, a confirmation the user actually reads, and sometimes a question back to the user mid-task (MCP elicitation). We never let a model write directly.
Whether the system has a clean API at all. A SaaS tool with a documented API is the easy case. An in-house tool, a legacy ERP or a desktop product with no web API (QuickBooks Desktop, for example) needs an adapter first, and that is often most of the work.

The cost after launch

The build is a one-time cost; the server is not a one-time thing.

The specification moves. MCP was revised in March, June and November 2025 and again in July 2026, each time touching how servers sign users in (the changes, revision by revision). The AI clients follow the newest version, so a server nobody maintains eventually stops connecting.
Your systems change. A new field, a renamed endpoint, a new permission. Each one is a tool to update and re-test.
Security is ongoing. Prompt injection through tool results, over-broad scopes and leaked tokens are real (how we lock our servers down). Someone has to watch the logs.

That is what the monthly running cost pays for. If a quote has no line for it, ask who will do it, and at what price.

When not to build one

An official server already does the job. Salesforce, QuickBooks Online, GitHub and many SaaS tools ship their own. Start there (when Intuit's connector is enough), and build only for the actions it lacks.
It is one read-only tool for you alone. A local server over one API is a short project with the official SDKs; our guide builds one in Go. Pay someone when it has to serve a team, sign people in or write.
You are not sure what the AI should do yet. Then the money goes into scoping, not code. The same logic is in what an AI agent costs.

How Much Does It Cost to Build an MCP Server? Our Real Prices (2026)

The cost guides for MCP servers quote ranges from a few thousand dollars to six figures. These are the prices we actually quote, what moves them, what it costs to keep a server running, and when you should not build one at all.