What is an MCP server? How AI assistants connect to your business, explained.
AI assistants are starting to do things, not just answer. An MCP server is how you let them work with your business: looking up what you allow and doing what you approve. Here's how it works, in plain English.
The simplest way to explain it
MCP stands for Model Context Protocol. It's an open standard, introduced by Anthropic in 2024 and now supported by Claude, ChatGPT, and a long list of developer tools, for connecting AI assistants to outside systems and data.
Think of it as one standard plug instead of a custom cable for every device. Before MCP, connecting an assistant to your booking system meant building something that only worked with that one assistant. With MCP, you build one server, and any assistant that speaks the standard can use it.
The server is the piece on your side. It tells an assistant what it can look up and what it can do, in plain language the assistant reads, and then does those things when asked. What it doesn't offer, the assistant can't do.
What an MCP server can do for a business
Four jobs, from sharing information to taking real steps.
Answer questions from the source
Your services, prices, hours, and service area, straight from your own data. An assistant stops guessing from a page it scraped months ago.
Take the next step
Check availability, book a consultation, or request a quote, once the person the assistant is helping says yes.
Put your team's AI to work
Pull a report, look up a customer, or draft a follow-up from your real systems, without anyone copying and pasting between tabs.
Connect what you already run
A CRM, a scheduling tool, invoicing, an internal database. Each becomes a few narrow, clearly described tools any MCP-capable assistant can use.
Public servers and private ones
Most businesses start with one of two kinds, and the rules that keep them safe are different.
Public: on your website
- Anyone's assistant can connect, with no login
- Shares what your website already shows
- Takes the same steps a visitor could, like booking a call
- Rate limited, and discoverable through a server card
upforge.io runs one. Its tools are listed further down this page.
Private: for your team
- Behind a token or a login, for named people or assistants
- Can do more, like send invoices or update records
- Risky actions default to a dry run or wait for a person
- Every action is traceable to who asked for it
We built one for a lawn care client that sends invoices and re-plans routes. Sends default to a dry run, and there are no delete tools.
A real one you can try: ours
upforge.io runs a public MCP server with 17 tools. Any assistant that supports remote MCP servers can connect to it; in Claude, add it as a custom connector. It needs no login, because it only shares what this website already shows and takes the same steps a visitor could. Agents can also find it on their own through our server card.
https://upforge.io/api/mcpServer cardLooks things up (11)
get_review_requirementsRead project review requirementsget_review_statusRead a private project reviewget_company_overviewGet company overviewget_servicesGet service catalogget_pricingGet pricing contextmatch_capabilityCheck whether Upforge can help with a specific needget_case_studiesGet portfolio case studies with outcomescheck_service_areaCheck whether Upforge serves a locationget_industriesGet industries servedsearch_faqSearch Upforge's FAQfind_consultation_timesFind open consultation times
Takes a step, with your OK (6)
prepare_review_uploadPrepare private project archive uploadfinalize_review_uploadFreeze a project archivesubmit_project_reviewSubmit a project for a static fit checkrequest_site_auditRequest a free website auditbook_consultationBook a consultationrequest_consultationRequest a consultation or quote
When an assistant books a call through it, the booking records which assistant made it and that the person agreed to be booked, and the person gets the same confirmation email, with links to cancel or reschedule, as someone who booked on the website.
How we build one
The code is the easy part. The decisions below are what make a server useful and safe.
- Pick the jobsStart from what people actually ask for. A handful of well-named tools beats a hundred that mirror every button in your software.
- Separate looking from doingTools that read never change anything. Tools that act are their own thing, each with limits written down.
- Decide who approves whatSome actions can run when asked. Messages to customers and anything involving money should default to a dry run or a person's sign-off. Some actions shouldn't exist at all.
- Give every agent a badgeEach assistant that connects gets its own identity, never a shared key, so every record says who did it and who approved it.
- Test with real assistantsTool descriptions are what an assistant reads to decide what to call. We test them in the assistants your customers and team actually use, and rewrite until the right tool gets picked.
- Publish it and watch itA public server gets a server card so agents can find it, a rate limit, and logs you can read. A private one gets revocable access.
What can go wrong, and how to keep it from happening
An assistant that can act on your business is useful for the same reason it needs limits. These are the three risks worth planning for.
An assistant doing more than you meant
Keep tools narrow and actions explicit. If a tool can't delete a record, no amount of clever prompting will make it.
Prompt injection
Text hidden in an email, a document, or a web page can try to give an assistant orders. Assume it will happen, and keep anything that moves money or reaches a customer behind a person.
Leaked or shared keys
One key per agent, stored on the server, never in a browser. When an assistant is retired, its access goes with it.
Where it fits with the rest of your site
An MCP server works best on top of a website that already answers questions clearly. The same structure that helps Google and AI answers understand your business, covered in what SEO is and our AI visibility work, is what an MCP server hands to assistants directly.
We build MCP servers as part of AI and automation, alongside the chat, automation, and insight tools on that page.
MCP servers, Answered
It's a small program that lets AI assistants connect to your business. It tells an assistant what it can look up and what it can do, and then does those things when asked. MCP, the Model Context Protocol, is the open standard assistants like Claude and ChatGPT use to talk to servers like it.
An API is built for developers who write code against it. An MCP server is built for AI assistants: every tool has a plain-language description the assistant reads to decide when to use it, and the server enforces what an assistant may and may not do. Often the best MCP server is a thin, careful layer over the API you already have.
Claude and ChatGPT both support it, along with a growing list of AI coding tools and agent platforms. Support for connecting your own server varies by product and plan, so we test yours in the assistants your customers and team actually use.
A public server should only share what your website already shows and only take the steps a visitor could take, like booking a call. Keep it rate limited, log what it does, and put anything that sends messages or moves money behind a person. That's how we build ours.
A chatbot is a conversation you host on your own site. An MCP server has no conversation of its own; it lets the assistants people already use, like ChatGPT or Claude, work with your business directly. Many businesses end up with both.
A server that only shares information is a smaller job than one that takes actions in your systems. Deciding who approves what, and testing with real assistants, is most of the work on the second kind. You get a timeline and a fixed price after a scoping call.
We quote it like any AI and automation project: a fixed price after a scoping call. The biggest factors are how many systems it connects to and how many of its tools take actions rather than look things up.
Want your business ready for AI assistants?
Tell us what you'd want an assistant to look up or do, and we'll tell you what the server would take, with a fixed price after the call.