Service · Updated
MCP Server Development.
Connect your software to any AI assistant once.
An MCP server is a small service that exposes your software's data and actions to AI assistants through the open Model Context Protocol. Build it once and Claude, ChatGPT, Cursor or your own agent can use the same tools, with the permissions you set. We design, build, secure and host MCP servers for business systems.
By George Shvaya, founder of Webvello · Published
Key takeaways
- MCP (Model Context Protocol) is an open standard, introduced by Anthropic in November 2024, for connecting AI assistants to tools and data.
- An MCP server exposes three things: tools (actions), resources (data to read) and prompts (reusable instructions).
- One MCP server can serve many AI clients, so you stop rebuilding the same integration for each assistant.
- The MCP specification defines OAuth-based authorization for remote servers, so each user can sign in with their own permissions.
- We built and run BalanceMCP, a hosted MCP server that lets an AI assistant work in a double-entry ledger.
What is an MCP server?
An MCP server is a program that speaks the Model Context Protocol and wraps one system, such as your CRM, a database or a set of internal APIs. The AI assistant is the client. It asks the server what it can do, then calls those tools when a task needs them. For the full background, read our MCP guide.
| Primitive | What it is | Example for a contractor |
|---|---|---|
| Tools | Actions the assistant can call | create_lead, find_open_slots, draft_estimate |
| Resources | Data the assistant can read | Price book, service-area list, job history |
| Prompts | Reusable instruction templates | "Summarize this site visit in our estimate format" |
What do businesses build MCP servers for?
CRM and job software
Reporting
Books and finance
Marketing data
Built by Webvello: BalanceMCP
Have a system your team keeps copying data out of?
Tell us which one. We will sketch the tools an MCP server would expose and what it would take to build.
How do you keep an MCP server secure?
- Per-user authorization. Remote servers follow the specification's OAuth-based authorization, so each person acts with their own permissions.
- Least privilege. Read-only tools by default. Write tools are narrow, named for what they do, and validated server-side.
- Human approval. Tools that send, delete, pay or change prices require a confirmation step in the client.
- Prompt-injection awareness. Text pulled from emails, forms or web pages is treated as data, never as instructions to run a tool.
- Audit logs. Every tool call is logged with the user, the input and the result.
- Rate limits and kill switch. Limits per user and a single setting to disable writes.
Local or remote: which kind do you need?
| Local (stdio) | Remote (HTTP) | |
|---|---|---|
| Runs on | The user’s own computer | A server you or we host |
| Best for | One technical user, developer tools | A team, or customers of a product |
| Sign-in | Uses local credentials | OAuth, per user |
| Updates | Each user updates their copy | Deploy once for everyone |
How we build one
- 1
Tool design
List the questions and actions people actually need, then name and describe each tool so the model picks the right one. - 2
Build with the official SDK
TypeScript or Python SDKs, typed inputs, server-side validation. - 3
Auth and permissions
OAuth for remote servers, scoped keys to your system, read-only first. - 4
Test with real clients
We test in Claude and other MCP clients on real tasks, and fix where the model misuses a tool. - 5
Host and monitor
Deploy, log, set limits and hand over the run-book.
Want to try it yourself first? Our walkthrough on how to build an MCP server covers the basics. If your team is adopting Claude Code, our Claude Code consulting covers MCP setup for developers.
Related services and resources
Pages that go deeper on the pieces this one touches.
Common questions
Straight answers to what owners ask us first.
What is an MCP server?
An MCP server is a program that exposes a system’s data and actions to AI assistants using the open Model Context Protocol. It offers tools (actions), resources (readable data) and prompts (instruction templates). Any MCP-compatible client, such as Claude or other assistants that support the protocol, can then use it.
Which AI assistants can use an MCP server?
Claude (desktop, web and Claude Code) supports MCP, and support has spread to other assistants, IDEs and agent frameworks. Client support for specific features varies, so we test the server in the clients your team actually uses.
Does our data leave our systems?
Only the data a tool returns for a given request is sent to the AI assistant. The MCP server decides what each tool can read, and it can run inside your own infrastructure.
How long does it take to build an MCP server?
A focused server with a handful of read tools and one or two write tools typically takes two to three weeks, including authorization and testing. Larger systems are built in stages.
Is MCP the same as an API?
No. An API is how software talks to software. MCP sits on top: it describes your API’s actions in a way AI assistants understand, adds a standard way to sign in, and works across many assistants without custom code for each.
Build the connection once
We will scope the tools, the permissions and the hosting, and show you a working prototype early.