Building a Smarter CRO Workflow with MCPs: How We Built Nucleus on Claude and Used Convert’s MCP
The intersection of AI, Model Context Protocols, and conversion rate optimization is reshaping how growth teams experiment, iterate, and win.
If you’ve been paying attention to the claude ai cro tooling space lately, you’ve probably heard the phrase “MCP” thrown around more times than you can count. But here’s the thing: most of the conversation is still theoretical. People are talking about what MCPs could do, not what they’re actually doing in real workflows.
We’re not here for the theory. We built Nucleus, an AI-powered CRO intelligence layer on top of claude ai cro, and we connected it directly to Convert’s MCP (Model Context Protocol). What followed was one of the most surprisingly productive engineering and growth experiments we’ve run in years.
This is the full story: why we did it, how it works, and what it actually changed about the way our team runs conversion rate optimization.
What Is an MCP, and Why Should CRO Teams Care?
Before we get into the specifics, let’s level-set on what a Model Context Protocol actually is because the marketing fluff around this term is genuinely exhausting.
An MCP is essentially a standardized interface that lets AI models like Claude interact directly with external tools, APIs, and data sources. Think of it as a plugin system, but one that’s deeply integrated into the model’s reasoning process rather than bolted on as an afterthought.
For CRO teams, this is a massive deal.
Traditionally, running a solid A/B testing workflow looked something like this: pull experiment data from your testing platform, export it to a spreadsheet, run some analysis, write your hypothesis in a doc, create a ticket in Jira, brief the developer, wait for implementation, then circle back to your testing platform to set everything up again. It’s a fragmented, context-switching nightmare.
MCPs collapse that fragmentation. When your AI assistant has direct, native access to your experimentation platform through a protocol like Convert’s MCP, it doesn’t just know your data it can act on it in real time, within the same conversation context.
Why We Built Nucleus, and Why We Built It on Claude
We didn’t set out to build a product. Nucleus started as an internal tool to help our growth team move faster without drowning in dashboards.
The problem we kept running into was simple but stubborn: our team had great instincts but not enough time to turn raw data into prioritized, well-reasoned experiments. The gap between “we should probably test this” and “here’s a fully briefed experiment with a clear hypothesis, target segment, and success metrics” was eating up hours every week.
We evaluated several large language models for the core reasoning layer. We landed on Claude for a few specific reasons that mattered for this use case.
Nuanced analytical reasoning. claude ai cro isn’t just about statistical significance. It’s about understanding the behavioral context why a user dropped off, what friction looks like at scale, where the data tells one story and the heuristic tells another. Claude handles this kind of layered, multi-variable reasoning better than models that lean too heavily on pattern matching.
Instruction-following without hallucination drift. In growth work, precision matters. If you ask an AI to draft a hypothesis based on a specific funnel stage, you need it to stay inside the guardrails of that context. Claude’s tendency to ask clarifying questions rather than filling gaps with confident-sounding guesses turned out to be exactly what we needed.
Extended context window. A full CRO workflow involves a lot of context: historical experiment results, current test hypotheses, audience segment definitions, business goals, and platform constraints. Claude’s context handling meant we weren’t constantly re-injecting state into every prompt.
MCP ecosystem readiness. Anthropic’s architecture for MCPs was already mature enough for production use by the time we started building. That was a meaningful practical differentiator.
Connecting Convert’s MCP: What It Actually Unlocked
Convert is a serious A/B testing platform used by enterprise and mid-market teams who care about statistical rigor and experimentation governance. They’ve built out a Model Context Protocol that exposes their core functionality experiments, goals, audiences, and reports directly to MCP-compatible AI clients.
Integrating Convert’s MCP into Nucleus changed three things in our workflow immediately.
1. Real-Time Experiment Awareness
Before the MCP integration, getting Claude to reason about our current testing landscape required us to manually paste in experiment data, status updates, and goal definitions. That’s not just slow — it introduces human error. You forget a running test. You paste stale data. You lose context between sessions.
With Convert’s MCP active, Nucleus has live awareness of every running experiment. It can see experiment names, variants, target audiences, goals, and real-time performance data without any copy-paste involved. When we ask Nucleus to generate a new test hypothesis, it can automatically cross-reference what’s already in flight, flag conflicts or overlaps, and factor in existing results without being prompted to do so.
This sounds like a small convenience. It is not. It’s the difference between an AI assistant that needs to be fed information and one that already knows what you’re working with.
2. Structured Experiment Briefing at AI Speed
One of the most time-consuming parts of our claude ai cro workflow used to be writing experiment briefs. A good brief needs a clear hypothesis, a defined control and variant, measurable success criteria, a minimum detectable effect estimate, and audience segmentation logic. Writing all of that from scratch, every time, is tedious and inconsistent.
Nucleus now handles this end-to-end. Given a conversion problem say, a drop-off on the checkout confirmation page it will:
- Pull relevant data from Convert to understand existing tests and baseline performance
- Draft a structured hypothesis grounded in behavioral logic
- Suggest variant directions with supporting rationale
- Define success metrics tied to Convert’s goal structure
- Flag recommended audience segments with estimated traffic volumes
What used to take 45 minutes of fragmented work now takes about four minutes of conversation. And because the brief is generated in context with live Convert data, it’s immediately actionable rather than requiring an additional translation step.
3. Post-Test Analysis with Narrative Context
This is the one that surprised us most. Post-test analysis is usually where insights go to die. You run a test, it ends, someone writes a quick summary, it gets filed in a Confluence doc that no one reads again, and the learnings evaporate.
With Nucleus connected to Convert’s MCP, post-test analysis becomes a genuine conversation. We can ask things like: “Why do you think the challenger variant underperformed on mobile despite winning on desktop?” and get a reasoned response that references the actual segment breakdown from Convert’s data, not just a generic answer about responsive design.
More importantly, Nucleus can surface connective tissue between experiments. It will notice if a pattern from a test six months ago is relevant to something we’re planning now, because it has access to the full experiment history through the MCP. That kind of longitudinal analysis used to require a dedicated analyst and a lot of calendar time. Now it’s a conversation.
The Technical Architecture (Without the PhD)
We’re not going to go deep on infrastructure here, that’s a separate post but the high-level architecture is worth describing because it explains why the whole thing works as smoothly as it does.
Nucleus is built as an MCP client. It connects to Claude via Anthropic’s API using the standard messages endpoint, with MCP server configurations passed in the request body. Convert’s MCP runs as a URL-based server, which means Nucleus can invoke it dynamically within the same reasoning context as a user conversation.
The practical effect is that when a team member opens Nucleus and starts a conversation, Claude isn’t just processing text, it has live tools available. It can call Convert’s MCP to fetch experiment data, then reason over that data, then suggest actions, all within a single coherent thread.
We also built a lightweight state layer that persists session context across conversations. This means when a growth manager comes back the next day to continue planning a test they discussed yesterday, Nucleus picks up the thread without needing a full re-briefing.
One architectural decision we’re genuinely glad we made: we kept the MCP integration modular. Convert is one MCP server in our setup, but the architecture makes it straightforward to add other analytics platforms, CMS systems, and customer data platforms. The pattern scales cleanly.
What Changed in Our Actual CRO Workflow
Let’s be concrete about the before and after, because that’s what actually matters.
Before Nucleus + Convert MCP:
The typical path from “we should test something” to “test is live” involves at minimum five separate tools, three handoffs, and usually a week of calendar time. Experiment documentation lived in scattered docs, post-test learnings rarely fed back into future planning, and the team’s best strategic thinking was constantly getting buried under execution overhead.
After Nucleus + Convert MCP:
The workflow is now almost entirely conversation-driven. A growth manager describes a problem or opportunity in natural language. Nucleus pulls relevant data from Convert, drafts a hypothesis, suggests a brief, and when approved, the brief goes directly to implementation with all the Convert configuration already specified. Post-test, the analysis feeds back into Nucleus’s context so future experiments are informed by the full history.
The qualitative shift is harder to quantify but possibly more important: the team is thinking more strategically because they’re spending less time on operational translation work. When you’re not copying data between tools and reformatting briefs, you have mental space for the harder question: why does the user behave this way, and what would actually change it?
Lessons Learned Building on MCPs for CRO
If you’re considering building something similar whether an internal tool or a product here are the honest lessons from our experience.
Start with a workflow that already has clear structure. CRO was a good fit for MCP-driven AI because the workflow has defined stages, documented artifacts, and measurable outputs. The AI didn’t need to invent the process; it just needed to accelerate and connect the existing steps.
MCP quality matters enormously. Convert’s MCP is well-built. The schema is clean, the data is reliable, and the available actions map sensibly to real CRO tasks. We’ve experimented with other MCPs where the schema was poorly documented or the data was inconsistent. The AI reasoning degrades fast when the underlying protocol is unreliable.
AI is not a replacement for expertise. Nucleus is significantly better at certain things than a human analyst speed, consistency, pattern recognition across large experiment histories. But it doesn’t replace the judgment of someone who deeply understands a specific business, customer base, or product. The best results come from treating Nucleus as a highly capable collaborator, not an autonomous decision-maker.
Explainability matters for team adoption. One of our early challenges was getting the growth team to trust AI-generated hypotheses. The turning point was making Nucleus show its reasoning not just the hypothesis, but why it was suggesting that hypothesis based on specific data points. Transparency drove trust faster than accuracy alone.
The Bigger Picture: Where CRO + AI Is Heading
The integration of Model Context Protocols into claude ai cro workflows is still early. Most teams haven’t touched it yet. But the direction is clear.
The future of conversion rate optimization isn’t about having more data or more testing capacity in isolation, it’s about having an intelligent layer that connects your data, your experimentation platform, your customer insights, and your strategic priorities into a coherent, responsive workflow.
MCPs are the connective tissue that make that possible. Platforms like Convert are building toward this future by exposing their functionality through protocols that AI can actually reason over, not just retrieve from.
Teams that figure out how to build this layer either by adopting tools like Nucleus or by building their own are going to have a meaningful, compounding advantage in their ability to learn and optimize faster than competitors.
That’s not hype. That’s just what happens when the bottleneck shifts from data access to decision quality, and the tools finally catch up.
Final Thoughts
Building Nucleus on claude ai cro and integrating Convert’s MCP wasn’t a moonshot. It was a pragmatic decision to stop accepting the fragmentation in our CRO workflow and build a better path through it.
The result is a team that experiments more thoughtfully, documents more consistently, and learns faster from every test. And honestly, it’s more interesting work because when you remove the operational overhead, what’s left is the genuinely hard, genuinely rewarding problem of understanding why people do what they do and what would actually change their behavior.
If you’re running CRO at any meaningful scale and you haven’t started exploring MCP-driven workflows, start there. Pick one integration Convert’s MCP is a solid first choice and see what happens when your AI assistant actually knows what you’re testing.
The gap between “AI assistant” and “AI collaborator” is narrower than you think. It just takes the right protocol to close it.
AEO | GEO | Claude ai cro
