UXArmy MCP server with Claude for Usability Testing: Can Claude Run a Study End to End?

Yes, with a human in the loop. I tested Claude MCP with UXArmy for usability testing end to end. Connected through the UXArmy MCP server, Claude turned my study plan notes file into an unmoderated usability study, set up recruiting from UXArmy’s participant panel, waited for my “launch it”, and, on a finished study, turned the results into a one-slide executive summary. I made the research calls; Claude did the setup and the synthesis.

Tested by Nick Lim, Product Growth, UXArmy · October 2026

What is an MCP server for user research?

Many research tools that shipped an MCP server this year use it only to read: search past studies, pull quotes, summarise transcripts. That’s useful, but it stops at the research you already have. It isn’t AI usability testing yet. I wanted to know whether an assistant could run an end-to-end usability study from scratch, so I tried it on the UXArmy MCP server before we launched it. This post is the honest log, including the parts that didn’t work.

If you just want to connect it, the setup guide covers Claude, ChatGPT, Gemini CLI and Cursor. The UXArmy MCP server page has the overview.

How do you set up Claude MCP with UXArmy for usability testing?

I picked a task I’d just done myself: booking a rental car on a travel site. I wrote the study plan as a plain notes file on my laptop (goal, audience, tasks and follow-up questions) and connected two things to Claude:

1. UXArmy: in Claude, go to Settings > Connectors, search for “UXArmy” and add it. You sign in on UXArmy’s own authorisation page, so no password or token ever goes into the chat.
2. Google Slides, through Claude’s own connector, for the hand-off at the end.

Then I asked: “Set up this study in UXArmy.”

A real example of Claude MCP with UXArmy: can Claude build a usability study from your notes?

Yes. Here’s what happened, step by step.

1. It reads the plan where it lives. Claude opened the notes file on my computer. No copy-pasting into the chat.

2. It told me what it couldn’t do, and what to do instead. Website usability studies can’t be created from chat yet (surveys can). Claude said so, and suggested I create an empty website study in UXArmy and let it fill it in.

3. It filled in the whole study. Scenario, screening questions, 6 tasks with the start URL on each, labelled rating scales (“1 = Very difficult … 7 = Very easy”) and open questions. It used my wording, and it didn’t invent success criteria or URLs I hadn’t given it.

Does Claude fix a study that’s too long?

Yes, and it didn’t get defensive. This was the part that surprised me most.

My study guide was too long. Claude had built every question in it, which came to 54 tasks and questions, so I asked whether that really fit a 30-minute session. Its answer started with “No, I’m not sure”, then broke the study down: 45–55 minutes in total, with 41 open-text answers as the main problem, since many repeated what the think-aloud already captures. I gave it a target of 25–30 minutes. It cut the study to 22 items by moving the follow-up questions into think-aloud prompts and adding a quick difficulty rating after each task.

It also flagged two setup risks before launch:

• Voice recording was off. That would have meant no think-aloud audio at all.
• Each task restarted on the homepage, which would have broken the booking flow between tasks.

💡 Give it a time budget: tell Claude how long the session should be (“fit this into 25–30 minutes”). It will trim the plan to fit and tell you what it moved or cut.

Can Claude recruit participants and launch the study?

Yes, from UXArmy’s own participant panel. I asked for 5 participants in Singapore, aged 25–50, on a computer, and Claude set the panel targeting. When I asked for narrower criteria that the panel filters don’t cover, it drafted them as screener questions instead.

Then it stopped. Nothing launched, and no credits were spent, until I typed “launch it”.

How do you analyze usability test results with Claude?

Setting up and launching a study is half the job, and many AI user research examples stop there. The other half is making sense of results and getting them in front of the people who decide. The car-rental study was still collecting responses, so I pointed Claude at a finished one: a desktop study in which 8 participants compared five travel insurers’ websites and then bought a policy on a client’s site.

I asked three plain questions: which account is this, what was the last study published, and what did it find. Here’s what came back.

1. It checked whose account it was in. It confirmed the signed-in user and my role before touching anything, and pointed out that the email it was signed in with wasn’t the one I’d typed.

2. It found the right study without IDs. I asked for “the last study published”. It listed every study on the account, picked the only one that had gone live, and said it was going by creation date because the connector doesn’t return a launch date.

3. It summarised the whole study. 8 of 13 responses were in, the median session was about 23 minutes, and every participant reached every task. It covered first impressions of the five sites (8 seconds each), a 1–5 likelihood-to-buy ranking where a client’s site came top at 3.75, median task times, what people liked and disliked in the purchase flow in their own words, and device preference: 7 of 8 would rather buy insurance on a laptop than a phone. 

4. It told me how far to trust the results. Alongside key takeaways and quick wins, it listed four caveats to raise before showing the data to stakeholders. The biggest finding suggested that with 8 participants that were all women, the findings are directional and not statistically conclusive.

5. It highlighted the finding that mattered. One participant only found out after choosing a plan that a key cover was a paid add-on, not part of the base plan, and it pushed her into a much pricier tier. Claude ranked that as the most serious finding, ahead of the slowest task: comparing plans took a median 3 min 41 s, and one person took over 10 minutes.

6. It checked what more research would cost. When I asked it to recruit more participants, it checked the balance first: 15 response credits and no panel credits. It worked out what the 5 missing panel responses would need (145 panel credits plus 10 more response credits), offered to invite my own participants instead, and warned that relaunching would also publish the unlaunched edits. It can’t buy credits, so it told me where to buy them in UXArmy.

Can Claude turn research findings into Google Slides?

Yes, through Claude’s Google Slides connector. I asked for one slide for a CEO with a short attention span. It built it with a headline (buyers trust the site most, but surprise pricing at checkout puts that lead at risk), three numbers (3.75/5 purchase intent and #1 of 5 sites; 5 of 8 called the site professional or trustworthy; 5 of 8 hit a friction point while buying), what works vs what hurts, one next step and the source study. It sized the text so nothing overflowed, then said it couldn’t see the rendered slide and asked me to check it.

The same hand-off works with whatever else you’ve connected to Claude, such as Slack, Notion, Jira or Linear. UXArmy doesn’t touch those tools; Claude’s own connectors do. I only tested Slides.

What can’t it do yet?

Being straight about the limits, as of October 2026:

• Website and prototype studies start in UXArmy. Create an empty study first; Claude fills it in and launches it. Surveys can be created entirely from chat.
• Some settings stay in the app. Voice recording, for example, is switched on in UXArmy, and credits are bought there too: Claude can check your balance but can’t buy credits.
• ChatGPT write actions need a Business, Enterprise or Edu plan. On ChatGPT Pro the connection is read-only.
• AI summaries need enough speech. When a participant barely talked, the summary read more into the session than the transcript supports. Check the transcript for short sessions.
• You’re still the researcher. Claude drafts, trims and summarises fast, but the time limit, the audience and the decision to launch were mine.

What did I learn?

Two decisions from me (the time limit and “launch it”) and Claude did the rest of the setup. What it saved me wasn’t only time. It caught a missing voice setting I’d have found only after the first sessions came back silent, and it showed exactly why my study was too long, and how to fix it, before any participant had to sit through it.

If your team already lives in Claude or ChatGPT, research no longer has to be a separate tab. The study, the participants and the summary can all come out of the same conversation.

Run your next usability study from Claude

Connect the UXArmy MCP server and go from notes to real participants in one conversation.

Sign up for free →

FAQ