MCP server development
In your customer’s AI assistant
Connect selected product features to an assistant such as Claude or ChatGPT. Customers connect their account before using the actions you make available.
Let customers use your product through AI.
For software companies with customer accounts and an API. We build MCP connections so customers can check orders, change bookings or find invoices through Claude or ChatGPT. Each connection uses their own login and your product’s permissions.
Connecting the tools you use to run your business? See business integrations.
Your customers know what they need. Connect the next step.
A customer should not need to know how your API works to use a feature through an assistant. We turn a useful product task into a connection that can find the right record, ask for missing information and return a clear result.
- “Where is my order?”
- Retrieve information from the customer’s account, including the status and details your product makes available.
- “Move my appointment to Friday.”
- Find an eligible time, ask the customer to confirm and return the appointment your system actually saved.
- “Find the invoice for this project.”
- Locate an account document and return an authorized link, without exposing another customer’s information.
The request follows your product’s rules.
Your customer asks their own assistant. The connection checks the account and your rules before your API reads or changes anything.
Example
- Customer:
Where’s my Harbor Supply order 1042? If it hasn’t shipped, ask them to use the side entrance.
- Assistant:
Harbor Supply · Read order1
Order 1042 is packed and ships Thursday. I can change its delivery note.
Update delivery note on order 1042?
- Now
- Leave at reception.
- New
- Use the side entrance.
- Assistant:
Harbor Supply · Update order2
Done. Harbor Supply saved the note: “Use the side entrance.”
- Customer:
Do the same for order 1038.
- Assistant:
Harbor Supply · Read order3
Order 1038 left the warehouse yesterday, so Harbor Supply won’t accept a change. Their support team can contact the carrier for you.
Only this customer’s orders
The customer signed in to Harbor Supply from their assistant. The connection reads orders from their account and nothing else.
No change without a yes
A change waits for the customer to approve it. The assistant then reports the note Harbor Supply saved.
Your product’s rules still apply
Shipped orders cannot be edited in Harbor Supply, so they cannot be edited through the assistant either.
What the connection can and cannot do.
These are implementation decisions for your team’s review. The protocol supplies a way to call tools; we build and test the checks around the tasks you choose.
- Use the customer’s own account
- Customers authorize access to their account. Each tool checks the customer and the permission required for that request, including a denied or expired connection.
- Separate reading from changing
- A lookup can return permitted information. Each write has its own allowed fields and confirmation step, with a result from your system before the assistant says the change is complete.
- Handle an incomplete request
- We test missing inputs, rejected actions and provider failures. The handoff identifies useful errors and cases that need your team, plus the logging and data-retention settings your team needs.
- Keep operator identity separate from customer access
- Web Bot Auth identifies an agent operator. OAuth grants access on a customer’s behalf. Operator identity does not grant permission to read that customer’s private records.
Technical references: Web Bot Auth and MCP authorization. Authentication, account mapping and permission checks remain part of the implementation.
What your team receives.
Start with a review of one customer task, its API requirements and the clients to test. You receive a written scope and acceptance checks. A build delivers source code, tool definitions, sign-in and permission tests, failure tests, and operating instructions. Optional care covers compatibility checks and updates as your API or clients change.
How we workStart with where your customers work.
The right connection depends on the task and the assistant or browser your customers want to use. We check those requirements with you before choosing the implementation.
MCP or WebMCP: read the buyer’s guideWebMCP implementation
On your website
Make suitable website actions available to a compatible browser agent. The customer can still use your ordinary forms and interface.
Direct API integration
In an agent you control
Connect your own assistant to the product features it needs. This can be a simpler fit when you do not need a shared protocol.
We confirm compatibility for the selected clients. WebMCP requires a compatible browser; a connection does not make your product available in every assistant. Technical references: MCP tools and Chrome’s WebMCP introduction.
Before we connect your product
How does MCP work with our existing REST API?
An MCP server gives an assistant tools that call your existing REST API. We shape those tools around customer tasks and check inputs and permissions. Your API still enforces its own business rules.
Can you work from our current API or OpenAPI specification?
Yes. An OpenAPI specification shows us the operations and data your product offers. We pick what the customer task needs and test it in each assistant you choose. Missing API operations are quoted before the build.
Which assistants can you support?
Claude and ChatGPT. We test the tasks in each one you choose, and one codebase can serve both. Extra interfaces and directory submissions are priced separately, and the directory operator decides approval.
How do customer sign-in and permissions work?
Customers connect their own accounts. We build access checks on your permission rules and test sign-in and expired access, even requests for another customer’s data. The assistant asks, and your system decides whether that customer may act.
Can we begin with read-only access and add actions later?
Yes. Start with lookups, such as finding an order or invoice. Changes to records come later, each with its own confirmation step. The connection offers only the tools you choose.
When is WebMCP a fit?
When the task happens on your website. WebMCP lets browser agents take actions on your pages, while a remote MCP server connects an assistant to your service from anywhere. We pick the one that suits the task and test it in the browser or assistant you use.
What happens when the protocol, our API or an assistant changes?
Your handoff lists what the connection depends on and how to rerun the task checks. A monthly fee can cover reviews and updates to existing tools. New actions or assistants are quoted separately.
Could our own engineering team build it?
Yes, if they know your API and the assistants well. We can also review their approach or build one part, such as the permission checks, with the split of work written down before we start.
Who hosts and owns it?
You own the code and the accounts it runs in. We include operating instructions so you can maintain the integration yourself or have another team take over.
What do you need from our team to start?
A customer task and a short description of your product. API documentation helps but is optional for the first call. The build needs documentation, a test environment and someone who can confirm product rules.
What should your customers be able to do?
Bring one customer task and we will work through the connection your product needs.