What MCP is
The Model Context Protocol (MCP) is an open standard that lets an AI assistant use outside services in a uniform way. A service publishes a list of tools. Each tool has a name, a description, the inputs it takes and the shape of what it returns. The assistant reads the list, picks a tool when the conversation calls for it, and gets a structured answer back.
The drawing shows the planned design. The MCP server is not available yet.
- The assistant reads the list of tools and picks one when the conversation calls for it.
- Read-only tools come first: capabilities, the design, the check report.
- Tools that write: uploads and quotes.
- The tool that commits money can only prepare an order. You pay on the site.
- The site answers in a structured shape, with exact numbers.
For you the effect is simple. Instead of the assistant guessing how a website works, the website tells the assistant exactly what it can do. The assistant can then do things like "check this board file" or "show me a price" in a way that is exact and repeatable.
Why a factory offers tools instead of a web form
A web form is made for human eyes. An assistant using one has to look at screenshots or scrape the page, and it breaks when the page changes.
A tool list is made for software. It says what each step needs and what comes back. For a circuit board order that matters, because the steps are precise: upload files, read the check report, change an option, read the price, confirm. Each step has numbers that must be exact, for example a hole size in micrometres or a price in a currency. Structured answers keep them exact.
Tools also make safety easier. A tool can say plainly that it spends money, and an assistant can be built to ask you before it uses such a tool. A web form gives no such signal.
What exists today and what is planned
Today. There is a REST API for the upload, file check and quote flow. It works without an account through a guest session. It has a public capability document, a report of 25 checks, a price preview and a kept quote. Ordering and payment are not open yet. The API quickstart shows the calls.
Planned, not available. The MCP server, API keys, sign-in for agents and ordering through an agent are planned and not available. We give no dates. The rest of this article describes the plan.
The safety model
The design is decided. It rests on four rules.
- An agent always works for a signed-in person. Agents have no accounts of their own. You sign in with your own account, and you allow the agent to act for you with named permissions. You can take the permission back.
- The agent can prepare an order. It cannot pay by itself. When an agent asks to place an order, the result is an order waiting for payment and a link to our site. You open the link, see the exact order and total, and decide. The link works once, is bound to your account, and stops working if the quote or the total changes.
- An engineer confirms every order. Whoever placed it, an engineer checks the files and the order before production starts.
- Read-only tools come first. Tools that only read, such as capabilities, the design and the check report, come before tools that write. Tools that commit money come last.
- You sign in with your own account and allow the agent to act for you.
- The agent prepares an order. It cannot pay by itself.
- The site shows the exact order and total behind a link.
- An engineer confirms every order before production.
- The amber path is the approval link: it goes to you, not to the agent.
There is deliberately no tool that changes your payment methods, your addresses or the limits of the account. Those exist only in the browser, where you are.
We also treat text from your files as data. File names, text inside design files and BOM descriptions come back in fields of limited length, never as instructions. Text in a file cannot change what the agent is allowed to do.
The planned tools
These are the tools in the design. They are a plan, and names and details may change.
| Group | Tools |
|---|---|
| Read what we can make | capabilities: options, limits, countries |
| Upload and read a design | create an upload session, get a design with its recognised layers and size |
| Check | validate a design, get the validation report |
| Quote | create a quote, update its options, get a quote |
| Order | place an order, get an order, list orders, answer an engineering query, cancel an order before production |
| Addresses | list addresses |
Placing an order returns the approval link described above. Cancelling is possible only before release to production.
What this means when it arrives
Say you ask an assistant to order twenty boards. It reads the capabilities, uploads your zip, reads the check report and tells you what it fixed. It creates a quote and shows you the price. It then prepares the order and gives you a link. You open it, check the total and the address, and pay. An engineer confirms the files, and production starts.
At no point does the agent hold your payment details. At no point can it spend without you. If its answer looks wrong, you see it in the quote before any money moves.
What MCP does not change
MCP is a way to reach the same service. It adds no new powers. An agent using the MCP server meets the same file check, the same limits, the same prices and the same engineer review as a person using the website. The permissions you give an agent are narrower than your own. A tool can only do what its permission allows, and you can remove the permission at any time.
What you can do now
You do not need to wait. Today you can use the same file check and price preview through the API, on your own or through an assistant that can call web services. If you work with an assistant, give it our PCB design guide for AI assistants. If you want to know how to work with an assistant in general, read Ordering boards with an AI assistant.
When the MCP server is open, this page will say so.