MCP server for your product catalog: let AI agents quote

Spokbee exposes your product catalog as an MCP server, so an AI agent can find a product, check a configuration against your engineering rules, price it and create a quote or a checkout link. The agent supplies the selections. Your configuration and pricing engines evaluate them, exactly as they do for a buyer on your website configurator.

An MCP server for a product catalog is the missing piece between "an assistant knows we exist" and "an assistant can buy from us". This guide explains what the Model Context Protocol is, which tools Spokbee gives a connected agent, what a buying conversation looks like step by step, how to switch agent access on, and how Spokbee keeps every agent-built configuration valid and correctly priced.

What is an MCP server for a product catalog?

An MCP server for a product catalog is a service that publishes your products, options, rules and prices as tools an AI assistant can call. The Model Context Protocol (MCP) is the open standard AI applications use to connect to external systems. With Spokbee the server is already built: publish a product and the tools exist.

The protocol's maintainers describe MCP as "an open-source standard for connecting AI applications to external systems" and compare it to a USB-C port: one standard connector instead of a custom integration per assistant. An AI application such as Claude or ChatGPT acts as the MCP host and opens a client for each server it connects to. A server exposes tools the model can execute, resources that provide context, and prompts that structure interactions. Remote servers use Streamable HTTP with bearer tokens or API keys for authorization. Claude, ChatGPT, Visual Studio Code and Cursor all support the protocol, according to the MCP introduction and architecture overview (checked October 2026).

A PDF catalog answers "what do you sell". It cannot answer "can I have the 48-inch cabinet in graphite with six shelves, and what does it cost". That second question is the one a buyer's assistant has to resolve before anything moves toward a purchase, and it depends on rules and prices only your configurator knows. Spokbee draws the line clearly on its own pages: AI product discovery helps an assistant understand what you sell, and agentic selling gives it actions to take next: read configuration options, validate selections, calculate a price and prepare a quote or checkout link.

What tools does Spokbee give an AI agent?

Spokbee gives a connected agent four MCP tools: find_products to discover published products, configure_and_price to check options and calculate a price, quote_product to create a configured quote, and checkout to prepare a buyer checkout link. A configure call returns validity, completeness, constraint violations, available options and a price breakdown.

The tool names and the response fields above come from the agentic selling page, which also shows an example configure_and_price call. Two design decisions matter more than the tool list. First, the website configurator and the agent interface use the same configuration and pricing engines, so the product knowledge you maintain once is useful wherever the buying conversation starts. Second, the tools an agent can see mirror the permissions you granted, so a catalog-only key never exposes checkout.

Developers who want to build their own integration rather than connect an assistant get the same actions over REST. The integrations page describes the REST APIs for working with configuration, pricing and quotes from your own application, webhooks that send events such as a new quote to an endpoint you manage, and the MCP tools themselves.

How does an AI agent configure and quote a product?

The agent reads the catalog, sends the buyer's selections to configure_and_price, and gets back a verdict: valid or not, what is missing, which options remain and the price. When the configuration is valid it calls quote_product, and Spokbee creates the quote with product, selections and calculated price travelling together.

Spokbee's agentic selling page walks through this with a sample workshop cabinet. The data is illustrative and no orders are placed, but the shape of the conversation is exactly what a connected assistant does.

Behind step 1, the assistant makes a tool call shaped like this. Product slugs and selection IDs come from your own catalog.

{
  "tool": "configure_and_price",
  "arguments": {
    "product": "workshop-cabinet",
    "selections": {
      "finish": "graphite",
      "width": "48",
      "shelves": "4"
    }
  }
}

Step 2 is the part that separates a configurator from a chatbot. The six-shelf request is not refused with a generic error. Spokbee returns the violated rule and the options that remain valid, so the assistant can propose a supported configuration in its next sentence instead of guessing. Step 3 only exists because you enabled it: quotes validate the configuration, and checkout uses your connected payment setup.

When the assistant calls quote_product, the result is a real Spokbee quote. The quotes documentation describes what that means: a branded PDF built from your template, with a table of every parameter and its selected value, a line-item price breakdown, buyer details and your terms, stored in the workspace Quotes dashboard and tracked through New, Viewed, Accepted and Rejected. A quote an agent creates sits in the same dashboard as a quote from your website, with the same lifecycle.

How do you connect an AI agent to Spokbee?

You need three things: a published Spokbee product with its configuration rules and pricing, a way for agents to reach it, which is either public agent access or a scoped agent key, and a compatible MCP client. Checkout additionally needs product-level enablement and a connected payment account.

That list comes from the "What do we need to get started?" answer on the agentic selling page. In practice it breaks into five steps.

Step 1 decides how good the agent's answers are. The guided selling documentation on the AI features page notes that an AI helper is most effective when parameters and options carry rich descriptions, and the same holds for an external agent reading your catalog. Write the option descriptions a knowledgeable salesperson would give. Encode the rules your engineers apply by hand. The constraints documentation covers six rule types: requires, excludes, conditional range, arithmetic, count limit and combination rules. A rule you write once applies to every website buyer and every agent.

Steps 2 and 3 are your access controls. Public agent access lets any buyer's assistant reach your published products the way anyone can open your public configurator, with no key and no account. Scoped agent keys are for integrations you authorize: each key gets exactly the permissions it needs, with rate limits and the ability to revoke access. Spokbee's workspace documentation on agent keys lists the permissions as catalog, configure, quote, checkout and export, and recommends starting with catalog and configure, adding quote and checkout only when the agent needs to complete transactions. Agent keys are prefixed spka_, shown once at creation, and can be rotated at any time, which immediately invalidates the old key.

Step 4 is a configuration entry, not a project. Remote-capable clients connect to Spokbee's hosted MCP endpoint over Streamable HTTP, presenting the agent key as a bearer token. For local desktop clients, Spokbee's workspace documentation describes a package that runs over standard input and output and forwards to the same API with your key. Either way, every call is evaluated on Spokbee's servers.

Step 5 keeps money where it belongs. The commerce documentation explains that Spokbee creates a Stripe Checkout Session for each order, that you retain full control of your Stripe account including refunds, disputes and payouts, and that Spokbee never holds funds. Checkout is enabled per product, so an agent can quote your whole catalog while only buying the lines you have chosen to sell that way.

Once agents are connected, agent analytics tracks requests, configuration activity, quotes and sessions, so you can see which assistants are active and what they are configuring.

Is the MCP server for public configurators or private ones?

Both. The MCP server exposes the products you have published, and you decide who can reach them. Turn on public agent access and any buyer's assistant can configure and quote them, exactly as anyone can open your public configurator. Leave it off and only agents holding a scoped key you issued get in, such as a dealer's procurement assistant or your own sales bot.

A product that is not published is invisible to every agent, so a configurator you are still building stays private by default. Beyond that, the workspace documentation on agent keys describes two doors into the same catalog.

Public access is the choice for products you already sell from a public configurator. Keys are the choice for everything else: a dealer network, a distributor's procurement system, or an internal assistant your reps use. Because anonymous traffic never sees per-key pricing rules, one catalog serves public buyers and authorized dealers at the same time without exposing a dealer's price to the public. Checkout behaves the same at both doors: it works only for products you have cart-enabled.

Your sales reps also have a home of their own today. The Sales Portal is a separate branded app where reps sign in to build proposals inside your configurators, with roles for admins, managers, reps and ops, so a key-based assistant and a human rep can work the same catalog and the same rules.

How does Spokbee keep agent orders valid and prices right?

Every configuration an agent submits is evaluated by the same constraint engine and pricing engine that run your website configurator. The agent never sets a price and never bypasses a rule. An invalid combination comes back as a constraint violation with the options that would fix it, so the agent revises instead of guessing.

This answers the question every engineering manager asks first: "What stops an AI from selling something we cannot build?" The buying assistant supplies selections; your configuration and pricing engines evaluate them. The agentic selling page states this directly, and the plan table on the AI features page lists Agent Commerce as a workspace capability on the Business and Enterprise plans.

Three layers do the work.

• Constraints. Requires and excludes rules cover most products: an outdoor finish requires aluminium or stainless steel; a glass material excludes the largest size. Conditional ranges narrow a dimension when another option is chosen, arithmetic rules validate expressions such as width times height, count limits cap selections of a type, and combination rules forbid or require specific multi-value combinations. All six are documented on the constraints page.
• Pricing. The pricing documentation covers base prices, surcharges and discounts attached to specific option values, formulas for anything that scales with a dimension, and volume tiers. An agent receives the resulting breakdown. It cannot edit it.
• Commercial controls. Quotes validate the configuration before they are created. Checkout is per product and runs through your Stripe account. Anonymous traffic through public agent access never sees negotiated per-key pricing rules, so a public assistant and an authorized dealer integration can share one catalog without sharing one price list.

Why open your catalog to AI buyers now?

Because buyers are already asking assistants. Gartner's March 2026 survey of 646 B2B buyers found that 67% prefer a rep-free buying experience and 45% used AI tools during a recent purchase. An assistant that can only read a PDF gives a vague answer. One connected to Spokbee gives a valid, priced configuration.

The figures are from the Gartner press release Gartner Sales Survey Finds 67% of B2B Buyers Prefer a Rep-Free Experience, based on a survey conducted from August through September 2025 (checked October 2026). The same research notes that buyers still want seller input when they need to judge whether a product fits their company. For a configurable product, that judgement is exactly what a rules-checked, priced configuration delivers, on demand, inside the assistant the buyer already uses.

Compare the two ways a complex product gets quoted.

The customer results Spokbee publishes come from the website configurator channel, and they show what the same engines do when the friction is removed. Stor-Loc replaced manual calculations and 2D drawings with visual configuration and automated quoting, and routine quotes now take half the time. Spokbee reduced PL Custom's time to quote on custom emergency vehicles from two years to a few days, with each configuration also delivering 2D drawings for engineering. An MCP server puts that same configure-and-quote capability in front of a buyer who arrives through an assistant rather than a browser, using the rules and prices you already maintain.

Sources

• What is the Model Context Protocol? and Architecture overview, modelcontextprotocol.io, checked October 2026.
• Gartner Sales Survey Finds 67% of B2B Buyers Prefer a Rep-Free Experience, Gartner, March 2026.
• Spokbee: Agentic selling, AI product discovery, Integrations, AI features, Quotes, Commerce and checkout, Constraints, Pricing, and the workspace documentation on agent keys, checked October 2026.