A custom MCP server does not have a list price, because the work is not the protocol — it is the tools behind it. Two servers that both “connect Claude to our CRM” can differ tenfold in effort depending on how many questions they answer, how good the platform’s API is, how strict the security requirements are, where it has to run, and who keeps it working when the platform changes. This guide explains each factor so you can read a quote intelligently, and gives you the questions that shrink one.
What you are actually paying for
The MCP protocol itself is open and the scaffolding of a server is a solved problem. The cost is in five places: designing and building the tools (the named actions the assistant can take), connecting them safely to your platform, building the identity and permission layer, deploying and hosting the result, and maintaining it as the platform, the model clients and your business change. If a quote does not break down along those lines, ask for one that does. Our MCP server development page describes how Inwizards approaches each part; this post is about what moves the number.
Factor 1: How many tools, and how narrow
Each tool is a small piece of software: input validation, a call to the platform, a shaped response, tests. A read-only server with five tools — find a customer, list their orders, check stock, get an invoice status, list this week’s deliveries — is a compact build. A server with thirty tools across sales, inventory, accounting and HR, with write actions and confirmations, is a different project. Counter-intuitively, narrow tools are cheaper to build and safer: a generic “query anything” tool is quick to write but expensive to secure, test and support. Start by listing the ten questions your team asks the system most; that list usually becomes the first tool set.
Factor 2: The platform’s API
This is the factor buyers underestimate most. Odoo has a clean, complete API and a predictable data model, so tools map onto it quickly. Zoho CRM and HubSpot have mature REST APIs with good documentation. Salesforce is rich but layered (objects, custom fields, sharing rules, flows) and needs care to respect. SAP varies enormously by version and module, and often needs an integration layer already present in the landscape. The WhatsApp Business API adds messaging rules (24-hour windows, templates) that the tools must enforce. Amazon’s SP-API has role approvals and rate limits that shape what a tool can do. The same five tools can be a short build on one platform and a long one on another purely because of the API underneath.
Factor 3: Identity, permissions and security scope
A server that runs under one service account with read-only access to non-sensitive records is the simplest case. A server that must act as each individual user, honour the platform’s own permission model, mask personal data in responses, require human confirmation on write actions, and produce audit logs a compliance team accepts, takes real design and testing time. Regulated industries should expect this layer to be a meaningful share of the build — and should not accept a quote that skips it. Our guide to MCP server security best practices lists exactly what belongs in this layer.
Factor 4: Where it runs
Hosting the server in a cloud environment you already use is straightforward. Running it on-premise — inside your network, sometimes with an open model behind it so no data leaves the building — adds infrastructure work: GPU sizing if a local model is involved, network access to the platform, secrets management, monitoring. For finance, healthcare, government and data-residency-bound companies this is usually the deciding factor and worth the cost; see our on-premise AI page for what that stack looks like.
Factor 5: Maintenance
Platforms change their APIs, model clients update how they call tools, your business adds fields and workflows, and someone has to keep the tool list current and the logs reviewed. A custom MCP server should be priced as a build plus an ongoing arrangement, whether that is a monthly maintenance fee or hours inside an existing support agreement. A one-off price with no maintenance path is a warning sign, not a bargain.
How long does a first build take?
Time follows the same factors as cost. A read-only server for one platform with a clean API and a short tool list is measured in weeks, and most of that is agreeing the tool list and testing against real records rather than writing code. Adding per-user identity, write actions with confirmations and a compliance-grade audit log extends it, and on-premise deployment adds infrastructure lead time on top. The parts that slip schedules are almost never the MCP layer: they are waiting for API credentials, discovering that the platform instance is customised in ways nobody documented, and security reviews that start after the build instead of before it. Bring the security team into the tool-list discussion on day one and the timeline shortens.
What makes a quote smaller
- Start read-only. Write tools, confirmations and rollback add work; add them after the read tools have proved their value.
- Bring the question list. Ten real questions from real users produce a tighter tool set than a vague “connect everything” brief.
- Use the platform’s permissions. Acting as the user, where the platform supports it, is less work than building a parallel permission model.
- Pick one system first. A server per system, built in sequence, is cheaper and safer than one server that spans four.
- Decide hosting early. Switching to on-premise mid-project costs more than choosing it up front.
Custom build, or the vendor’s official MCP server?
Salesforce, HubSpot and Dynamics 365 ship official MCP servers, and there are open-source Odoo MCP modules. For a standard setup with default objects and modest security needs, those may be enough, and the honest answer is to try them first. The custom route earns its cost when you need your own custom objects and fields, per-user permission enforcement, response filtering for sensitive data, a specific set of write actions with confirmations, on-premise deployment, or tools that span your customisations. We build the custom version; we also tell you when you do not need it.
How to get an accurate quote
Send the platform and version, the list of questions or actions you want, who will use it and what they may see, whether write actions are needed, the hosting constraint, and any compliance requirement. With those six inputs a competent team can give a fixed price for a first build rather than a range. Inwizards builds custom MCP servers for Odoo, Zoho CRM, SAP, the WhatsApp Business API, Salesforce, Dynamics 365, HubSpot and Amazon Seller Central this way, priced as fixed-scope builds with a maintenance path, from teams in the US, UAE and India. We deliberately do not publish a price here, because any number without those inputs would be a guess dressed up as a fact.