A custom MCP server that lets AI use your system on yourterms.
Your team already asks an AI assistant for help, but it cannot see your orders, bookings or records. A custom MCP server gives it a controlled way in: a short list of tools you choose, read-only first, with sign-in and a record of every call.
An MCP server opens a door for AI. Who else can walk in?
MCP is a young standard, and many servers were put online in a hurry. Two studies from 2025 counted how many were left open, and how many still depend on one fixed key.
1,862
MCP servers were found exposed to the internet in one scan by security researchers
What we doWe agree with you who should reach the server, and keep a server meant for your team off the public internet.
Five worries business owners bring to an MCP server developer, and what we do about each.
The problem
My team uses an AI assistant every day, but it cannot see our orders, bookings or records.
What we do
We turn the jobs people really ask for into a short list of tools. Each tool has a clear name, a plain description and a strict list of inputs, because that description is all the AI model has to go on.
No way in4 done for you
One tool for each real job
Descriptions a model can follow
Strict inputs, checked on every call
Tested in the AI apps your team uses
The problem
The server we have exposes one tool that can run anything, and the assistant uses it freely.
What we do
We replace wide tools with narrow ones and start with tools that only read. A tool that changes something is added later, with a limit on what it can touch. Most AI apps ask the user before a tool runs, but that is the app's choice, so the server enforces its own rules as well.
Too much power4 done for you
Read-only tools first
No tool that runs free-form commands
A limit on each action
A person confirms where you want it
The problem
A customer could write instructions into a ticket, and the assistant might follow them when it reads it.
What we do
This is called prompt injection, and no server can rule it out. We limit what it can cause. Tools return only the fields the job needs, in a fixed structure. Tools that send or change something stay separate from tools that read outside text, and we test with planted instructions.
Tool results4 done for you
Only the fields the job needs
Results in a fixed structure
Reading and acting kept apart
Tested with planted instructions
The problem
Everyone on the team connects with the same key, pasted into a settings file on each laptop.
What we do
For a server reached over the web, we use the sign-in flow the MCP standard describes, which is built on OAuth. Each person signs in with their own account, and the server only accepts access that was issued for this server.
Sign-in4 done for you
Each person signs in as themselves
No shared key in a settings file
Access issued for this server only
A permission for each tool
The problem
If the assistant changed a record last week, I have no way to find out who asked or what it did.
What we do
The server writes down every call: who was signed in, which tool ran, what went in and what came back. We also limit how often each tool can be called, so one confused agent cannot flood your system.
No record4 done for you
Every call written down
Tied to the person who signed in
A limit on calls for each tool
A log you can read and search
Three MCP servers. Three lessons.
Two of these are demonstrations by security researchers and one is a flaw that was reported and fixed. No theft was reported in any of them. None of them is our client.
01
Security research · July 2025
A support ticket that made an assistant leak a private table
What happened
Researchers at General Analysis built a test project on Supabase and connected the Cursor coding assistant to it through Supabase's MCP server. A support ticket carried hidden instructions. When the developer asked the assistant to look at the latest ticket, it read a private table of tokens and posted the contents into the ticket. The assistant was using a database role that could read everything.
What we do about it
We give each tool only the access its job needs, start read-only, and keep tools that read outside text apart from tools that write.
An example server that trusted its input, copied thousands of times
What happened
Trend Micro reported that Anthropic's example MCP server for SQLite databases put user input straight into a database command. The Register reported that the code had been forked or copied more than 5,000 times before it was archived, and that Anthropic said it would not be patched because the repository was archived. Trend Micro showed how the flaw could be used to plant instructions for an AI agent.
What we do about it
We check every input against a strict schema and never build a database command from raw text. We also read example code before we build on it.
One build setting that reached thousands of hosted servers
What happened
GitGuardian researchers found that Smithery, a service that hosts MCP servers, did not check a path in its build settings. That let them read a token from the build machine, and the token gave control of more than 3,000 hosted apps, most of them MCP servers. Smithery fixed it within days of the report in June 2025, and no evidence of misuse was found.
What we do about it
We host your server under your own account, and give the server and its build only the access they need.
Live websites we built for technology companies. They show our build quality, not MCP servers. The MCP server we can show you is the one on this website, at /mcp. Work for private clients is shared on a call, with the client's agreement.
Why choose Sun Software for MCP server development?
Getting a server to answer is quick. Deciding what it should be allowed to do is the work.
We have built one, and you can connect to it. This website runs a small MCP server at /mcp with three tools that only read: company details, a list of pages, and one that returns a page as Markdown. It needs no sign-in because it returns nothing the site does not already publish, and a server card tells AI agents where it is. It is a small server, and we have not yet delivered one for a client that we can show you. We build yours the same way: read-only first, then actions with limits, sign-in where the data is private, and a log of every call.
The Sun Software team
01
A server you can inspect today
Our own site has a read-only MCP server with three tools and a public server card.
02
We tell you when an API is enough
If your own software is the only thing calling the model, a direct API call is simpler.
03
Tools designed, not generated
Each tool is written for one job, with only the inputs and results that job needs.
04
Built for a standard that changes
MCP has had at least five dated revisions since late 2024. We record which ones your server speaks.
Great....very professional, delivered on time and exactly as I needed. Exceptional. I always find Sunsoft very easy to work with, always helpful and I'll continue using their services for the foreseeable future. No need to go anywhere else.
Client in United KingdomOctober 2026 · Fiverr review
“
Excellent experience from start to finish. Communication was great, responses were always prompt, and the work was completed quickly. They captured the exact look and feel I wanted on the first try and delivered excellent value for the money. I would gladly work with them again and highly recommend their services.
Client in GermanySeptember 2026 · Fiverr review
“
Choosing SunSoft was exactly the right decision. Their professionalism, extensive experience, and attention to detail have really impressed me so far. It was a complex project that exceeded my expectations! I look forward to continuing to work with SunSoft!
Client in CzechiaAugust 2026 · Fiverr review
Questions about MCP servers, answered plainly
What is an MCP server, in plain words?
It is a small program that sits in front of your system and offers a list of named tools, such as “find an order” or “list today’s bookings”. An AI app that supports MCP reads that list and calls a tool when a person’s request needs it. MCP itself is an open standard, published at modelcontextprotocol.io.
Do we need a custom MCP server if we already have an API?
Often the API is enough. If your own software calls a model and you control both ends, calling your API directly is simpler and we will say so. An MCP server earns its place when people want to use your system from AI apps you do not control, or when several assistants need the same tools.
Which AI assistants will be able to use it?
Any app that supports MCP. Support differs from app to app and from version to version, for example in whether an app can reach a server on the web and how it handles sign-in. We test with the apps your team uses and write down what worked.
Can an MCP server stop prompt injection?
No, and nobody can promise that. Text a tool returns can carry instructions, and a model may follow them. What a server can do is limit the damage: narrow tools, read-only by default, sign-in for each person, limits on actions and a log. We test with planted instructions and show you the results.
MCP keeps changing. Will my server stop working?
The standard is young and it does change. The July 2026 revision removed the opening handshake and the sessions that earlier revisions used. A server can be built to answer both older and newer AI apps. We record which revisions yours speaks, and we can keep it current under a care plan.
Have you built an MCP server yourselves?
Yes, a small one. This website has a read-only MCP server at /mcp with three tools: company details, a list of pages, and one that returns a page as Markdown, a plain text format. It needs no sign-in because it only returns what the site already publishes. We have not yet delivered an MCP server for a client that we can show you.
What decides the cost of an MCP server?
You get a written proposal once we know which jobs the server has to do and what it connects to. Three things decide it. The tools: How many tools the server offers, and how many of them change something instead of only reading. Your system: Whether your system already has a documented API we can call, or whether that has to be built first. Sign-in and hosting: Who may use the server, whether it is public or internal, and whether you already have a sign-in service it can use.
What happens after I contact you?
We reply within one business day. Then the work runs in 6 steps: Choose the jobs, Design the tools, Set sign-in and permissions, Build the read-only version, Test with real assistants, Add actions and release. Each step ends with something you can review.
1Talk About Your MCP Server
2We reply within one business day
3Discovery call
4Step one: Choose the jobs
Not sure you need an MCP server?
Tell us what you want an assistant to do in your system. We reply with questions, then a tool list you can approve.