Brisora MCP
What is Brisora MCP?
Brisora MCP is a remote procurement decision engine for AI clients. It lets MCP-compatible systems call deterministic tender qualification, commercial-policy evaluation, landed-cost analysis and sourcing comparison capabilities while keeping procurement decision logic outside the AI client.
The AI client decides when to call a tool and how to explain the answer. It does not produce the answer. Brisora MCP analyses and evaluates only: it cannot send, submit, order, pay, sign or approve anything.
Why MCP.
The Model Context Protocol is an open standard that lets an AI client call external tools. Offering Brisora through MCP means the engine can be used from Claude, OpenAI-powered agents and other MCP-compatible clients without adopting another workflow platform.
It also draws a clean line. A tool call has a defined input and a defined output, so the procurement rules stay on Brisora's side of that line and cannot be rewritten by a prompt.
A call travels one way and comes back unchanged.
The MCP layer is a thin adapter. It validates the request, passes it to Brisora's application services, and serializes the result. There is no model call anywhere on that path. The full layering is on the architecture page.
- A fresh server instance handles each request and is then discarded
- Nothing is stored between requests, or at all
- The only outside system contacted is the public TED website, to fetch the notice you name
Current tool surface
Four tools, all read-only analysis.
The surface is deliberately small and closed. Each tool is described in full on the page that owns its topic.
Qualify a public tender notice
Fetches one notice from TED and qualifies it under a named policy profile. Detail: tender qualification and its evidence.
Evaluate an action against a mandate
Checks a proposed commercial action against a mandate you supply. Detail: commercial-policy evaluation.
Calculate landed cost and margin
Exact arithmetic over the prices, components and exchange rates you provide. Detail: the landed-cost model.
Compare bill-of-materials options
Returns which options are valid, which need review, and the recommended selection. Detail: sourcing option comparison.
Every result includes its reason codes and rule trace. What that trace contains is explained under procurement decision provenance.
Division of responsibility
What belongs to the AI client, and what belongs to Brisora.
The AI client
- Understands the question and holds the conversation
- Chooses which tool to call and with which inputs
- Explains the result in plain language
- Cannot change a rule, skip a check or supply a missing fact
Brisora
- Retrieves and retains the evidence
- Applies qualification, policy and commercial rules
- Produces the outcome, reason codes and rule trace
- Keeps unknowns as unknowns
Demo data boundary
What the demo accepts and what the server cannot execute.
Brisora MCP has no tool that sends email, sends a request for quotation, submits a tender, accepts a quote, creates a purchase order, makes a payment, signs a contract, changes a mandate or approves an escalation. These capabilities are absent from the server, not disabled in it.
The demo has no user accounts and no tenant separation, so it does not support private or customer data. Do not send confidential tenders, customer documents, supplier-sensitive terms or credentials.
Protocol and transport
A standard remote MCP server.
Streamable HTTP
The current MCP transport for remote servers, built on the official MCP SDK. No provider-specific logic.
Stateless
No sessions. Each request is self-contained, and two identical requests produce identical results.
Bounded
Request size, argument size, execution time and request rate are capped. Errors are fixed codes without internal detail.
Demo connection model
How a client will connect.
Claude connects to remote MCP servers as custom connectors. OpenAI-powered agents attach a remote MCP server as a tool. Other MCP clients connect over Streamable HTTP in the same way.
Brisora MCP is being opened through a private demo program. Request access and we will provision a demo connection once your use case fits the current public-data boundary.
Live Claude and OpenAI client connections are being proven before broader demo distribution. The server is tested against the official MCP client; no public endpoint is published.