What custom MCP development costs, and how to scope it
A custom MCP server can be a small project or a large one. A simple local build is typically a low four-figure job, and a production server with user sign-in reaches five figures. Four things explain most of the gap: authorisation, the number of tools, hosting and the depth of security review. All figures on this page are indicative.
Indicative budget ranges
The ranges below are orientation, drawn from how such projects are usually sized, and they help you plan a budget. For a quote on your own scope, BlackCard prices each project after a free 15-minute call.
| Build type | What it includes | Indicative budget | Typical duration |
|---|---|---|---|
| Local server for a small team | A few read-only tools, an API key read from the environment, packaging and install notes | Low four figures in euros | One to three weeks |
| Remote server with sign-in | OAuth 2.1 through your identity provider, a handful of tools, hosting, tests in two or three clients | Low to mid five figures | Four to eight weeks |
| Server with write actions and review | Confirmation steps, audit log, per-user permissions and an independent security review | Mid five figures and up | Two to three months |
| Programme across many systems | Several servers, a gateway, a registry and shared rules for teams | Six figures | Several months |
Driver 1: authorisation
Sign-in is the largest single variable. A local server for one developer reads an API key from its environment and needs almost no authorisation work. A remote server used by many people is a different project. Under the specification it acts as an OAuth 2.1 resource server: it has to publish metadata that tells clients where to sign in, validate that each token was issued for it, and obtain separate credentials for the systems behind it instead of passing the user's token along.
Connecting that to your identity provider, mapping roles to scopes and testing the flow in each client can take as long as writing the tools.
Driver 2: number and quality of tools
Each tool needs a name, a description the model can act on, an input schema, error messages that help the model recover, and tests. Read-only tools are cheaper than tools that write, because a write needs confirmation logic, idempotency and an audit trail. As a rule of thumb, a build with one or two tools and a build with half a dozen integrations sit in different price brackets, and the second can cost about twice the first.
Driver 3: hosting and operations
- Local distribution costs little to run but needs packaging, installation instructions and updates on every user's machine.
- Self-hosted remote needs a deployment pipeline, monitoring, rate limits and someone on call.
- Platform-hosted moves that work to a hosting product in exchange for a subscription or usage fee.
- Ongoing change is a real line item: the specification is revised regularly, and the 2026-07-28 revision alone deprecated several features that older servers rely on.
Driver 4: security review
A server that can change records or read personal data should be reviewed before launch. The protocol's security guidance lists the main risks, among them confused deputy attacks, token passthrough and over-broad scopes. A review covers those, plus how the server behaves when a model is fed hostile text through a document or ticket.
In regulated sectors the review, the documentation and the sign-off meetings can cost more than the code. Ask for this to be priced as its own item so it is not quietly dropped.
How to scope a first project
A scope written this way fits on one page. It is the starting point BlackCard works from, because it settles the first useful version before any code is written. Bring it to a free 15-minute call, or let the team write it with you.
- Pick one system and one group of users.
- Write down five to eight tasks and mark which ones change data.
- Decide local or remote, and name the identity provider if remote.
- State which AI clients must work on day one.
- Ask for a fixed price on that scope and a day rate for changes.
Common questions
Why do quotes for the same server differ so much?
Mostly because developers assume different things about sign-in and review. One quote may cover a local server with an API key, another a remote server with OAuth, per-user permissions and a penetration test. Send every bidder the same written scope, including the clients to support and the data the server may change, and the spread narrows.
Is a platform cheaper than a custom build?
A platform can lower the starting cost if you have developers and a documented API, because it removes hosting and much of the authorisation work. The cost that remains is design: deciding which operations become tools and how they are described. That design work decides how well the server performs, and it is where a senior team such as BlackCard earns its fee.