UX audit with a real user, fixed by your coding agent
A real person tries one flow of your app on camera. Your coding agent turns the recording into a UX audit with timecoded fixes, within 24 hours.
- One real person, one flow
- Back within 24 hours
- Your agent fixes what you approve
In short: your coding agent sends one flow of your app to a real person who has never seen it. They try it on camera while thinking aloud. Within 24 hours the recording comes back to your agent, which writes a UX audit: a short summary and a table of fixes, each linked to the second in the video that proves it. Nothing in your code changes until you pick the fixes.
What you get
- A real person on camera. Screen, webcam and voice, plus a log of every click, while an in-house tester tries your flow for 15 to 30 minutes.
- A summary in plain words. Three to six sentences on how hard the task was, grounded in time spent, the worst delay or wrong clicks.
- Three to six proven fixes. Each with a linked timecode and a severity: conversion risk, friction or polish. A problem the recording cannot prove is left out.
- Fixes when you say so. The agent asks which fixes to make, then changes the code. Test again with a new person to check the result.
How it works
- Describe the flow. In your agent, one sentence is enough: Have a real user check out as a guest on staging. Or let
plan_usability_test_suitedraft the critical flows of your project. - A real person tries it.
create_usability_tasksends the flow to an in-house tester who has never seen your product. You see what the agent sends and approve it first. - The recording comes back. Within 24 hours the session is ready, and
get_usability_resultbrings it into your agent when you ask for the status. - Your agent writes the audit.
plan_jam_ux_auditmakes it prove each problem from the click log and the transcript, link the exact second of the video, and rank the fixes. - You choose what to fix. Reply with the numbers, such as Do 1 and 3, and the agent makes those changes.
Example audit
What an audit looks like for a guest checkout on a fictional online store. In a real audit every timecode opens the recording at that second.
The tester reached the confirmation page in 7:40, on the second try. Most of the time went to payment: after a card error at 4:12 the Pay button was out of view, and the tester scrolled for almost two minutes and asked whether the money had been taken. Finding the guest option and the delivery date cost another minute.
| # | Fix | Evidence | Severity |
|---|---|---|---|
| 1 | Keep the Pay button in view when the card form shows an error | 4:12, the error pushed Pay below the fold; the tester scrolled up twice and asked “did it take my money?” | Conversion risk |
| 2 | Put “Continue as guest” above the sign-in form | 0:48, the tester started creating an account, then went back | Friction |
| 3 | Show the delivery date in the cart | 2:31, the tester asked when the order would arrive and opened the FAQ | Polish |
The audit ends with a question: which of these should I fix?
Why a real person
- It shows what happened, not what might happen. A rule-based or AI review lists what could confuse someone. A real person shows where they actually hesitated, misread a label or gave up.
- Every fix has proof. The audit follows a strict order: clicks first, then what the tester said, then screenshots. Console and network logs only explain a problem the clicks already show.
- The analysis runs on your agent. Claude Code, Claude, Cursor or Codex writes the audit on your own subscription. Hallway Test runs no AI model.
- We never see your code. The agent sends the link to test and the task. That is all.
How it compares
| UX audit with a real user | Expert UX audit | AI review of screenshots | |
|---|---|---|---|
| Who looks | A person who has never seen the product | A UX specialist | A language model |
| Scope | One flow per test | Usually the whole product | What you show it |
| Evidence | Recording, transcript, click log | Expert judgment and principles | The model’s judgment |
| Turnaround | Within 24 hours | Depends on the scope | Minutes |
| Fixes | Your agent applies the ones you pick | A report for your team | Suggestions |
What it costs
The first test is free: one usability test and a 30-minute feedback call. After public launch a test costs $29, with no subscription. Your agent writes the audit on your own plan, so Hallway Test does not charge for it. See pricing, or read how the usability testing MCP works.
Mobile apps too
React Native and Expo apps work the same way: the tester installs your build from a link, such as TestFlight, and the Hallway Test SDK records the session. You can also record sessions of your own app and let your agent audit them, see mobile app recording.
Questions, answered
A usability test of one flow, written up as an audit. A person who has never seen your product tries the flow while thinking aloud, and the session is recorded. Your coding agent then reads the recording and lists the fixes, each proven by a moment in the video.
An expert audit judges the whole product against UX principles. This audit shows where a real person got stuck in one flow, with the recording as evidence. It is narrower, faster and cheaper, and the fixes land straight in your code through your agent. Many teams use both.
A coding agent that supports MCP, such as Claude Code, Claude, Cursor or Codex, and a link to the build you want tested: production, staging or a preview. If the flow needs an account, add test login details to the task.
No. Your agent plans the test inside your project and sends only the link to test and the task text, plus test login details if the flow needs them. You see what it sends and approve it first.
Your coding agent writes it, on your own subscription or API key. Hallway Test runs no AI model: its tools return the recording data and the rules for the audit, and your agent does the analysis.
The example on this page is the template: every audit comes back in the same shape, a short summary and a table of three to six fixes with timecodes and a severity. You can ask your agent to export it to a file or an issue tracker.