AI Agents Can Now Request Aquaculture Quotes and Run Engineering Calculators

For twenty years, B2B procurement software assumed the buyer was a person with a browser. That assumption is quietly breaking. A growing share of early-stage project work — sizing a recirculating system, checking whether a country can supply fingerlings, estimating a feed budget, drafting a specification — now happens inside an AI assistant before anyone visits a website at all.

So we did the obvious thing and made the platform callable. Alongside the human interface at fishmatchgroup.com, FishMatch Group runs two agent interfaces built on the Model Context Protocol (MCP), the emerging standard that lets an AI assistant discover and use external tools safely:

• A public, read-only server at /api/public/mcp. No account, no key. It exposes what is already published: sourcing categories, priority markets, the calculator catalog, live calculator execution and cost benchmarks.
• An authenticated server at /mcp. Once a user signs in, their assistant can create a sourcing project, attach the specification and follow its status — the same objects the sourcing desk works from.

The commercial point is not novelty. It is that a specification prepared by an agent is usually more complete than one typed into a contact form at 23:00, and a complete specification is the single strongest predictor of a fast, comparable quote round.

The interaction looks nothing like a chatbot answering from memory. A typical sequence for a buyer planning a 500-tonne tilapia RAS in Ghana runs like this:

• The agent calls list_markets to confirm Ghana is a served market and to retrieve its species and system profile.
• It calls list_sourcing_categories to see which equipment verticals can be sourced — filtration, oxygenation, hatchery, feed systems, processing, cold chain.
• It calls run_calculator with real parameters: target biomass, stocking density, temperature, feed rate. Back come tank volume, biofilter load, oxygen demand and aeration sizing, with the assumptions attached.
• It calls get_cost_guide for the species and system to place a CAPEX band and an OPEX profile around those numbers.
• It drafts the brief. If the user is signed in, the authenticated server turns that brief into an actual project; if not, the agent hands the user a link to request a quote with the work already done.

Every number in that chain is computed by our engine, not invented by a language model. That distinction is the entire value of tool access.

The public surface exposes the same engineering and economics tools human buyers use on the calculators hub:

Results are returned with their inputs and assumptions, so an assistant can show its work — and a reviewing engineer can challenge it. Sizing outputs are planning-grade starting points, not stamped engineering drawings; the detailed design is done by the supplier or consultant who wins the scope.

Preparation is only half the job. The part buyers actually care about is what happens after the brief exists. On FishMatch Group the flow is unchanged whether the brief came from a person or an assistant: one technical specification goes confidentially to the small set of manufacturers that genuinely build that scope, quotes come back against that identical specification, and the buyer receives a single comparable set.

What agent access changes is the clock. Clarification rounds — missing tonnage, undefined water source, no target harvest weight, no delivery incoterm — are the usual reason a quote round takes five weeks instead of two. An agent that has already run the calculators and filled the specification checklist removes most of them before the request is ever sent.

Opening a platform to machines raises an obvious question: what stops an agent from harvesting the supplier network? The answer is architectural rather than contractual. No tool on either endpoint returns a supplier name, website, email, phone number or address. The public server touches only published content; it has no path to projects, leads or the supplier database. The authenticated server is scoped to the signed-in user's own projects and nothing else.

Most of the current discussion about generative engine optimization stops at content formatting: answer-first paragraphs, FAQ schema, clean entities. Those matter, and we use them on every page. But the next layer is functional. When an assistant can call a tool instead of paraphrasing a paragraph, the site stops being a source of sentences and becomes a source of answers — computed, current and attributable.

For a buyer that means fewer wrong numbers in the planning phase. For us it means our data appears in AI answers in the form we published it, with the assumptions intact. We think that is where B2B discovery is heading in every capital-goods category, not just aquaculture: published prose for humans and crawlers, published functions for agents, and one confidential commercial process behind both.

Point any MCP-compatible assistant at https://fishmatchgroup.com/api/public/mcp and the tool list appears automatically — nothing to install, no key to request. To create and track real projects, sign in on the site first and connect the authenticated server at https://fishmatchgroup.com/mcp.

Prefer to stay in the browser? Everything the agents can reach is available to you directly: the project planner, the calculators, the priority market hubs and the financing center. The agent interface is a faster door into the same building.