I built an open-source tool that lets AI agents hire real humans to test your product

The problem

I use AI coding tools (Claude Code, Cursor) for almost everything. Development speed is incredible — what used to take a week now takes an afternoon.

But there's a gap nobody talks about: "it compiles" ≠ "users can use it".

I kept shipping features that worked perfectly in my terminal but confused real users. And manual QA was eating all the time I saved from AI coding. The irony was painful.

Traditional UX testing tools like UserTesting or Maze didn't help either — they're built for product managers reviewing dashboards, not for AI agents writing code.

What I built

human_test() is an open-source platform that closes the loop between AI coding and real user experience.

Tell your agent: "Test my app at localhost:3000, focus on the signup flow"

Then your agent calls human_test(), 5 real humans test your product with screen recording and audio narration, AI analyzes the recordings and generates a structured report, finds 3 critical issues, auto-generates fixes, and creates a PR.

You do nothing.

The full workflow

1. 
   Create a task — provide a URL (or description for mobile/desktop apps) and what to focus on


2. 
   Real humans test — testers claim the task, record their screen and microphone, go through a guided feedback flow (first impression, task steps, NPS rating)


3. 
   AI generates a report — extracts key frames from recordings, uses vision AI to analyze usability issues, aggregates everything into a structured report


4. 
   Auto-fix — if you provide a repo URL, it clones your code, generates file-level diffs, and creates a PR

The entire loop runs without you touching anything after the initial call.

What makes it different

The key insight: this is not a dashboard for humans to interpret. It's a structured API that AI agents can call, parse, and act on directly.

The report is designed for machines. Each issue has a severity tag like CRITICAL, MAJOR, or MINOR, plus three fields:

[CRITICAL] Signup button unresponsive on mobile

• 
  Evidence: 3/5 testers couldn't complete registration on iPhone


• 
  Impact: 60% of mobile users will abandon signup


• 
  Recommendation: Fix touch target size, minimum 44x44px

Your agent reads the severity, reads the Recommendation, and writes a targeted fix. No human interpretation needed.

Other differences from traditional UX tools:

• 
  Webhook-driven — async notifications when reports and code fixes are ready


• 
  Auto-PR — from usability issue to pull request, fully automated


• 
  Self-hostable — runs locally with SQLite, your data stays on your machine


• 
  Open source — MIT licensed

Try it

Option 1: Self-host (3 commands)

npm i -g humantest-app

humantest init

humantest start

Local SQLite database, zero external dependencies.

Option 2: AI agent skill (1 command)

npx skills add avivahe326/human-test-skill

Works with Claude Code, Cursor, Windsurf. Then just ask your agent in natural language: "Run a usability test on my checkout flow with 3 testers"

Option 3: Hosted version

Use human-test.work — zero setup.

## Tech stack

Next.js 16, Prisma, NextAuth, Tailwind CSS. Supports Anthropic and OpenAI for report generation. SQLite for local dev, MySQL for production.

Links

• 
  GitHub: github.com/avivahe326/humantest
  


• 
  Live: human-test.work
  

I'd love to hear from you:

• 
  If you use AI coding tools — how do you handle usability testing today? Do you just ship and hope for the best?


• 
  If you try human_test() — what's your first impression? What's missing?


• 
  If you've built similar tools — what did you learn?

Drop a comment below or open an issue on GitHub.