Custom MCP servers: your systems, inside Claude and ChatGPT.
I build MCP servers that let Claude, ChatGPT and other AI assistants read and work with your business systems, through tools named for the job and limited to what each person may see.
- Starts with
- Free 30-minute call
- Works with
- Claude, ChatGPT, any MCP client
- Access
- Read-only first, per person
- Based in
- Los Angeles, California
(01)What it is
The Model Context Protocol (MCP) is an open standard for connecting AI assistants to tools and data. An MCP server is the piece that sits next to your system and offers the assistant a short list of tools, such as “look up an order” or “run this report”, each with clear limits.
Once it’s connected, people ask in plain English inside the assistant they already use, and the answer comes from your own records rather than the model’s memory.
(02)Use theirs or build one
Some systems ship their own. Many needs don’t fit them.
| System | Official MCP option | When a custom server helps |
|---|---|---|
| NetSuite | Oracle’s NetSuite AI Connector Service: NetSuite provides the MCP server; it needs a role with the MCP Server Connection permission (the Administrator role can’t use it). | Your own saved searches as named tools, your business rules, or NetSuite answered together with other systems. |
| HubSpot | HubSpot’s remote MCP server, connected with OAuth, for CRM data. | HubSpot joined with your ERP or store, or actions shaped around your own pipeline. |
| QuickBooks Online | Intuit’s open-source MCP server, released as an early preview to run locally. | A hosted server your whole team can use, with each person’s access set on purpose. |
| Your own database or app | None: it’s yours. | Read-only views and the questions your team actually asks, as tools with clear names. |
| Several systems at once | Each vendor’s server covers only its own system. | One server that answers across your store, CRM and ERP, such as “which open orders belong to customers with overdue invoices?” |
(03)How it’s built
- 01Tools named for the job. “Find open orders for a customer” rather than a raw API call, so the assistant picks the right one and can’t wander.
- 02Read-only first. Each action it may take, such as creating a draft, is added on purpose after reading works well.
- 03Each person’s own access. People sign in with OAuth and the server acts with their permissions, never a shared admin key.
- 04Hosted where it fits. A remote server works from Claude and from ChatGPT where your plan offers custom connectors; a local one stays on one machine.
- 05Every call logged. What was asked, which tool ran, what it read or changed.
- 06Tested on real questions. A list of the questions your team asks, run before launch and after every change.
(04)Built this way
On my own product: Safe Streets Map.
Safe Streets Map turns California’s public crash records into maps and thousands of pages. I built an MCP server over its data with four read-only tools: find a city’s exact name, a city’s totals, the streets with the most crashes, and the crashes near a point. Ask Claude “which streets in Pasadena have the most serious crashes?” and the answer comes from the records, through tools that can read but never change them.
Swap crash records for your orders, invoices and inventory, and it’s the same build.
(05)When not to
Honest scope
If your vendor’s official MCP server already answers what your team asks, use it and skip the build. A custom server is for questions that cross systems, rules only your business has, or systems nobody has connected yet.
(07)Questions
What is an MCP server?
A small service that offers an AI assistant a set of tools and data through the Model Context Protocol, an open standard. Any assistant that supports MCP, including Claude and ChatGPT, can then use those tools: look up an order, read a report, create a draft. The server decides what each tool can see and do.
Do I need a custom MCP server if my software already has one?
Often not. If the vendor’s official server does what your team needs, use it. A custom server earns its place when you need several systems answered together, your own rules or saved searches as tools, tighter limits than the vendor’s, or a system with no server at all.
What’s the difference between an MCP server and an API integration?
An API integration moves data between systems on a schedule or trigger, with no one asking. An MCP server waits for an assistant to ask, then answers or acts within the tools it offers. Many businesses need both: the integration keeps systems in sync, and the MCP server lets people ask questions across them.
Does ChatGPT support MCP?
Yes, as custom connectors to remote MCP servers on the ChatGPT plans that offer them. Claude supports MCP in its desktop app, on claude.ai and in Claude Code. A server built to the protocol works with both.
Is it safe to connect AI to our systems?
It’s as safe as the server’s limits. Read-only tools, each person’s own permissions, drafts instead of final records, keys kept on the server and every call logged are how the risk stays small. What the model provider keeps is set by the plan you use, which we check first.
What does it cost?
A free 30-minute call first, then a fixed quote for a first server with a handful of tools against one system. I’m based in Los Angeles, California and work with businesses across the U.S.
(Next step)
Have work AI should be doing? Let’s map it out.
Start with a free 30-minute call: what you want connected, automated or found, and the smallest first step worth paying for.