MCP in procurement: connecting your own AI assistant
The Model Context Protocol (MCP) is an open protocol through which an AI assistant discovers and calls tools that another system offers. In procurement it lets the assistant a company already uses ask the procurement system what is bought, where it is delivered, from whom and at what price, and hand in a proposal for what to put to suppliers. A safe connection gives the assistant tools that read and one tool that proposes. The server computes every figure of a proposal itself, a person approves or dismisses it, and no assistant key can send a message, approve, award or change anything.
The Model Context Protocol (MCP) lets an AI assistant call tools that another system offers. For procurement this means that the assistant your staff already use can ask the procurement system what is bought, where, from whom and at what price, and can hand in a proposal. The connection is worth having when it ends there: the assistant reads and proposes, and people decide.
What the Model Context Protocol is
MCP is an open protocol that standardises how AI applications connect to external data and tools. Anthropic published it in 2024, and the specification and documentation are at modelcontextprotocol.io. It defines three roles.
- Host. The AI application a person works in, for example a chat application or a coding tool.
- Client. The part of the host that holds the connection to one server.
- Server. The system that offers something to the assistant.
A server can offer three kinds of thing: tools, which are functions the model may call, each with a name, a description and a schema for its input; resources, which are data the application can read; and prompts, which are prepared templates. Messages are JSON-RPC 2.0. A server runs locally and talks over standard input and output, or remotely over HTTP. A client opens with initialize, asks with tools/list what the server offers, and calls a tool with tools/call.
The practical effect is that a system writes one server, and every assistant that speaks MCP can use it without an integration of its own.
What it is good for in procurement
When someone in purchasing, finance or a plant puts a question to an assistant that has no connection, the answer comes from an export somebody pasted in, which is out of date the next day, or from the model's general knowledge. With a connection, the assistant fetches the answer from the system that holds the invoices:
- What did we spend on packaging in the last twelve months, by site?
- Who delivers cleaning supplies to our second plant, and at what prices?
- Which items have no agreed price and are bought from several suppliers?
- Put together a proposal for a request for quotation for these items for both sites.
The first three are reading. The background is the same data that a spend analysis rests on. The fourth is a proposal, and it is the furthest an assistant should go.
What the assistant may do
| Level | What it covers | Who answers for the result |
|---|---|---|
| Read | Spend overview, items with their prices and suppliers, delivery sites, suppliers on record, earlier requests and what was agreed | The server, which returns computed facts |
| Propose | Hand in a proposal: these items, for these sites, to these suppliers, with a reason | The server computes quantities, prices and amounts. A person approves or dismisses |
| Not available | Send a message to a supplier, approve, award, change data or settings | A person, in the procurement system itself |
How to design the server side
- One key per assistant. Scoped to read or to propose, revocable, shown once and stored only as a hash.
- Read tools return computed facts. Sums, averages and yearly quantities come from the server, so the assistant does not add anything up itself.
- Exactly one tool writes, and it writes a proposal. There is no tool that sends or approves, so no instruction can reach one.
- Take nothing numeric from the caller. The assistant names items, sites and suppliers. Quantities, today's prices and the yearly amount are computed from invoices.
- Check what is named. Items, sites and suppliers must exist in the organisation. A refusal comes back as a tool error with its reason.
- Check the words. A reason that states a figure no tool returned is replaced by a computed sentence.
- Log and limit. Every stored proposal is in the audit trail under the key's name, and requests are rate limited per key.
- Offer a key without amounts. Some assistants should see what is bought and where, and no prices.
What goes wrong
- Instructions hidden in data. An assistant reads item descriptions, supplier names and replies, and text of that kind can say "ignore your rules". The server cannot stop an assistant from being misled. It can make sure that a misled assistant has nothing harmful to call.
- Suppliers from the model's memory. An assistant proposes a supplier it "knows". Nobody can say which page that rests on. Accept suppliers on record, and leave research to a process that keeps the source.
- One key for everything. A shared key with the widest scope cannot be revoked for one assistant alone.
- Sums made in the chat. When the assistant adds figures up itself, the result is a model's figure. The article on why AI must not invent figures explains why that matters.
How Cureman does it
Cureman offers an MCP endpoint as part of the module sourcing and negotiation. It speaks the protocol over plain HTTP: one JSON-RPC request or a batch in, one JSON answer out, without a session and without a stream, with the methods initialize, ping, tools/list and tools/call. An organisation creates assistant keys in its settings. A key is shown once, only its hash is stored, it can be revoked, it has the scope read or propose, and it can be made blind to amounts. On a customer's own host name only that customer's keys are accepted.
The read tools return twelve months of spend by category, site and supplier, the groups of items that could be put to suppliers now, a single item with its suppliers, prices and agreements, the delivery sites, the suppliers on record and the earlier sourcing events. The one writing tool stores a proposal. It goes through the same function as a proposal of the built-in buyer: items, sites and suppliers must exist, quantities, today's prices and the yearly amount are computed from invoices for the chosen sites, and a reason that states a figure of its own is replaced. Web research for new suppliers is not offered over MCP, and a proposal takes suppliers on record only, because a supplier named by an outside assistant could not be tied to a page the system opened.
A stored proposal appears in the list of decisions without a worth, since nobody has quoted. Approved, it becomes a draft sourcing event, and from there every request for quotation is approved and sent by a person, as described in human approval for AI agents. The rules behind this are on the page about principles, and how it works describes the daily work around decisions.
Questions and answers
What is the Model Context Protocol (MCP)?
- MCP is an open protocol, first published by Anthropic in 2024, that standardises how an AI application connects to outside systems. A server offers tools, resources and prompts, and a client inside the AI application discovers them and calls them with JSON-RPC messages. One server then works with every assistant that speaks the protocol.
What can an assistant do in procurement over MCP?
- It can answer questions from the organisation's own purchasing data, such as spend by category, site and supplier, the price history of an item or the outcome of earlier requests for quotation. With a wider scope it can propose a request for quotation. It should not be able to send, approve, award or change anything.
Is it safe to connect an AI assistant to procurement data?
- It is as safe as the tools the server offers and the key the assistant holds. A key that can only read still shows data to the assistant and whoever operates it, so it should be scoped, revocable and, where needed, blind to amounts. A tool that proposes is safe when the server computes all figures itself and a person decides about every proposal.
Why should an assistant not send requests for quotation itself?
- A request to a supplier speaks for the organisation and opens a negotiation. An assistant reads text from many sources, and such text can contain instructions that the assistant follows by mistake. If the furthest it can reach is a proposal in a person's list, a mistake costs one dismissal.
How does Cureman offer MCP?
- Cureman has an MCP endpoint over HTTP as part of its sourcing module. An organisation creates assistant keys with the scope read or propose and can make a key blind to amounts. Every proposal is stored by the same function the built-in buyer uses, which computes quantities, prices and amounts from invoices, and no assistant key can send, approve, award or change anything.