Redmond built a production AI commerce agent in 10 weeks using Shopify's Storefront MCP

Redmond has been mining natural salt from an ancient deposit in central Utah since 1958. What started as two brothers pulling salt out of the ground has grown into a company with more than 10 brands, including Real Salt and Re-Lyte, the electrolyte hydration supplement that took off during the pandemic. The business runs on Shopify Plus.

When Redmond's managed AI customer service tool was discontinued, the team had a choice: pay for another vendor, or build their own. 

Phillip Hinson, an applied AI developer who had been building internal knowledge tools at Redmond, and Jeremiah Payne, the development team lead who had managed Redmond's Shopify infrastructure since 2018, decided to build. 

Using Shopify's Storefront MCP as the foundation, they built a production AI commerce agent in 10 weeks. It went live in February 2026 and is already handling customer conversations at scale.

Key results

• A production AI commerce agent built in 10 weeks by a two-person team
• Thousands of customer conversations handled monthly with high accuracy
• Full control over system prompt and response guardrails for product-specific messaging
• Minimal ongoing maintenance after initial stabilization
• Store consolidation completed internally. Payne used MCP to migrate historical customer and order data from three legacy Shopify stores, eliminating a paid third-party app.

The backstory: From 35 WordPress sites to Shopify Plus

Jeremiah Payne joined Redmond when the company had approximately 35 different websites, most of them running on separate WordPress and WooCommerce installations that were difficult to maintain at scale.

Payne had worked with Shopify at previous companies and knew what it could handle. He made the case internally to migrate to Shopify: it would eliminate the management and maintenance burden of constant upgrades, letting the team focus on selling. They started by migrating the Redmond Equine brand before moving on to Redmond Life.

The timing of the migration turned out to be fortunate. COVID accelerated ecommerce demand, and Re-Lyte launched into a hydration market that was gaining mainstream attention. Redmond already had an online presence and used the moment to grow rapidly on Shopify.

That was 2018. Seven years later, the Shopify operation has grown significantly across DTC and wholesale, and the team is now consolidating four separate Shopify stores into a single Shopify Plus store with Hydrogen-powered storefronts for each brand.

The challenge: Scaling CX without losing control

As Redmond grew, the team looked for ways to handle the increasing volume of routine product questions without scaling headcount proportionally. Redmond's deliberate hiring culture (multi-step interviews, cultural alignment checks) meant the team couldn't add CX capacity as fast as demand was growing.

"We are very careful in our hiring. Onboarding new people takes a really long time here," Payne says. "We were looking for ways to scale our customer experience without proportionally growing the team."

In early 2024, Payne began evaluating managed AI customer service tools. He found a solution that worked as a first point of contact for customers, handling product questions and basic discovery before routing to live agents when needed. Redmond became an early AI adopter, and customers adapted quickly to using a chatbot as their first interaction with the brand.

The tool confirmed the demand for AI-assisted customer service, but over time, it also exposed a structural limitation: the team didn't have sufficient control over the AI's responses or the data behind them.

The control gap

Redmond's customers routinely ask detailed questions about ingredients, product composition, and sourcing. For a natural products company, accuracy on these questions is critical — the team needed to control exactly how the AI communicated about their products.

At the time, managed AI chat tools generally didn't expose system prompt controls to merchants. For most use cases, that's fine, but for a company that needed precise guardrails around product-specific claims, the lack of control was a dealbreaker.

"With any managed tool, there's a trade-off between convenience and control," Hinson says. "We couldn't customize the system prompt or control the data sources, which made it difficult to build guardrails around the specific product questions we knew required careful answers."

When the managed AI service was discontinued, the team needed a quick path forward.

Payne did a cursory search for replacements. The market didn't offer anything that solved the control problem. The cost of the managed service gave them room to invest in an internal build.

"We were leaning towards building in-house anyway because of the control issue," Payne says. "AI development has made it easier for everyday people to do development work."

The solution: Building an AI ecommerce agent using Shopify's Storefront MCP

Payne had seen the Storefront MCP in a Shopify announcement and asked about it during a monthly check-in. He received additional documentation and brought it to Hinson.

Hinson had been building internal knowledge management tools using RAG (retrieval-augmented generation) pipelines and had been experimenting with MCP protocols. When he saw the Storefront MCP documentation, the architecture clicked.

"I felt like we could do it ourselves," Hinson says. "Shopify had released their Storefront MCP about five or six months earlier. Once I saw it, it felt like exactly the foundation we needed. The pieces were there to put it together."

How the system came together in 10 weeks

With the Storefront MCP as the foundation, the team was ready to move from setup to production. The GitHub commit history traces how the system came together over the next 10 weeks. Here's how it unfolded:

Initial setup: Hinson started with Shopify's MCP reference app, which provided the application scaffolding, authentication flows, and MCP client integration. He connected it to a Redmond development store and deployed to Azure with a CI/CD pipeline. Getting the MCP connected and responding in the development store took about a day.

"I was just so surprised how easy it was to get started and wire it up in the development store," Hinson says. "It's like a USB-C. It should work almost everywhere."

Content pipeline: Hinson built a sync that pulls content from all of Redmond's blogs and web pages across their sites, storing it in a PostgreSQL database. This provides the RAG layer — the brand-specific knowledge that the Storefront MCP's product data alone can't answer. Product-specific questions pull from this curated knowledge base, not from web scraping.

Semantic search: Hinson created text embeddings for all content using Azure OpenAI and stored them in a vector database for semantic retrieval. When a customer asks a detailed product question, the system retrieves the specific, approved knowledge base article rather than generating an answer from general training data. The conversational AI itself is powered by Anthropic's Claude, which interprets the customer's question, decides when to search, and composes the response.

Prompt caching: Hinson implemented Anthropic's prompt caching for the AI responses. Because the system prompt and tool definitions are large and stable, caching them avoids re-processing on every request. This cut input token costs by up to 10x on cache hits and reduced overall AI spend by roughly half.

Analytics and attribution: Hinson wired up token tracking, tool usage tracking, and a multi-tier attribution system that matches purchases to chat conversations at varying confidence levels, from direct cart tracking to time-window correlation.

Payne notes the team deliberately chose conservative attribution: "We're wanting to be a little bit more targeted to try and get a more honest perspective of what our tool is doing. We don't want data that makes it look like it's running better than it is."

Production hardening: Security implementation, PKCE authentication flow, log sanitization, and rate limiting through Azure App Insights.

Human handoff: Analytics showed customers consistently wanted the option to speak with a person. Hinson built a "Talk to a Human" button that provides a seamless transition from the AI chat to a live agent in HubSpot's chat widget.

Total build time: 10 weeks. Hinson handled the development work, while Payne provided direction, check-ins, and strategic decisions.

Learning from the previous tool

Before decommissioning the previous tool, the team exported conversation logs from the prior nine months. Analyzing what customers asked — and where answers fell short — gave them a clear picture of what the new system needed to handle well from day one.

The results: What changed after launch

The agent went live in February 2026, timed to launch the same day the previous service was discontinued. The transition was seamless, with no service interruption for customers.

Accuracy on product-specific questions

The product-specific questions that had been challenging for the managed AI tool are now handled with precision. 

When a customer asks a detailed question about a product, the agent provides a nuanced, accurate answer with a direct link to the relevant knowledge base article. The response is pulled from Redmond's curated content and runs through a system prompt that the team controls.

Every response regarding product composition, ingredient sourcing, and product claims follows the same pattern: curated source data, a controlled system prompt, and direct links to approved documentation.

Real-time product data without manual updates

The moment that captured the MCP's value came during launch week. Redmond released a new protein powder product the day after the agent went live, while Hinson was on vacation in Mexico with his family. 

A colleague messaged Hinson, congratulating him on quickly updating the agent's product knowledge for the new protein powder. Hinson hadn't done anything. The Storefront MCP had automatically pulled the new product from Shopify's Catalog, and the agent could answer questions about it immediately.

"To me, that's the real value of building on a live data protocol," Hinson says. "There's no sync to run, no content to manually update. The MCP connects directly to Shopify's Catalog in real time — the agent just knows about new products the moment they're published."

With the managed AI service, new products required manually initiating a content update. With the MCP, Catalog changes in Shopify are available to the agent in real time.

Low operating costs with full control

Anthropic API token costs have been minimal since launch. That does not include internal development time during the 10-week build. Since launch, maintenance has been minimal, following an initial stabilization period of bug fixes for authentication and attribution in the weeks after launch.

The previous managed service was a significant line item. The comparison isn't apples-to-apples — the managed service was an all-in subscription covering hosting, AI model, support, and updates. Redmond's custom agent separates those components: Azure for hosting, Anthropic for the AI model, and internal development time for the build and maintenance.

What Redmond traded was a managed subscription for full ownership of the system, the compliance controls, the data, and the ability to extend the agent in any direction.

"We didn't build this to save money. We built it because we needed control," Hinson says.

CX team refocused

The agent now handles product discovery, ingredient questions, and routine order inquiries, addressing the scaling challenge the CX team was facing. By automating repetitive, high-volume questions, the team can focus on interactions that require judgment, nuance, and customer care. A built-in handoff ensures customers can always connect with a real person when needed.

Faster, lower-cost store migration with MCP

The Storefront MCP's utility extended well beyond the chatbot. When Redmond began consolidating four Shopify stores into one, Payne evaluated third-party apps for migrating historical customer and order data. After looking at pricing, he realized he could do it himself using the Shopify MCP and Claude.

"Last week, I did that and all of our customer data from those other three stores historically has been moved over," Payne says. "And then I also built a middleware app that listens to a webhook: anytime a new customer is created or an order is paid for, it automatically migrates those over to the other store."

A migration project that would have required a paid app or professional services (scoped, priced, and waiting in a queue) was completed internally in a fraction of the time.

"The Shopify MCP has definitely given us a lot of tools and ability to do things faster and cheaper than what we were able to do before," Payne says.

What's next: Transactional AI, Hydrogen, and a unified multi-brand experience

With a production-ready agent in place and control over both the experience and underlying data, Redmond is now expanding on three fronts simultaneously.

Transactional agent capabilities

The current agent handles informational conversations. The next phase adds transactional workflows, starting with refund and cancellation processing for unfulfilled orders. Redmond's internal policy is already permissive on refunds, so the AI can handle the request end-to-end while asking follow-up questions to capture data on why customers are canceling.

Hydrogen-powered unified storefront

Redmond has been building Hydrogen storefronts for each brand for over a year. The development work is largely complete and running on a test store. Customers will be able to sign in to a single account across all Redmond brands, add items from different storefronts to the same cart, and check out once.

"Re-Lyte is our front door where everybody sees, everybody knows, and we want to invite them into the rest of the house," Payne says.

The launch is waiting on a subscription platform migration that moves subscription customers from three legacy stores to the unified platform without losing any active subscriptions.

The advice for other teams

Hinson and Payne both emphasize that the barrier to entry is lower than most teams assume.

"Having passion for this space is a must because things are changing so fast," Hinson says. "But the barrier is genuinely lower than most teams think. The hard part isn't the AI — it's knowing your product well enough to teach it."

Payne adds, "AI development has made it easier for everyday people to do development work. I don't think it would be a much bigger lift for most teams to take this on."

Shopify's developer infrastructure (including the Storefront MCP, GraphQL APIs, and Hydrogen framework) is designed for teams like Redmond's to build custom commerce experiences.