← All articles

Customer story

Coding agents write the code. Holmes checks it.

How a small team ships three web products and two apps, tested on every pull request, without writing a single test script.

A conversation with

Jens BogaertFounding Engineer, Maurice & Nora
Company
Students helping families and older people at home
Region
Flanders and Brussels
Stack
Supabase · Vercel · Expo
With Holmes since
April 2026
5Web products and apps under test
0Test scripts written by hand
Before mergeEvery pull request tested on its Vercel preview

The problem

Five surfaces, coding agents, and no one left to click through it all

Maurice & Nora matches screened students with families and older people who need help at home: groceries, company, a ride to an appointment. They operate across Flanders and Brussels.

Behind that service is more software than most teams their size look after: an admin tool, a client portal and a student portal in one monorepo, plus client and student apps built with React Native and Expo. A lot of the code starts with coding agents that pick up Linear tickets and open pull requests on their own.

Linear ticketCoding agent

Clicked through by hand

Before every release, if there was time

Admin

Unchecked

Client portal

Unchecked

Student portal

Unchecked

Client app

Unchecked

Student app

Unchecked

When agents and engineers ship this fast across five surfaces, nobody can click through everything before a release. Jens and his colleagues Klaas and Tiemen wanted testing that keeps up with how they build, without writing and maintaining a Playwright suite by hand.

What changed

First tests in one afternoon, described instead of scripted

Maurice & Nora started with Holmes in late April 2026. After the first call, the Holmes team opened a shared Slack channel and asked for blunt feedback. They got plenty.

The moment it clicked came early. Jens connected Holmes to Claude through MCP and used it to turn their biggest features into Holmes tests. Nobody wrote a script: they described what matters, and a browser agent runs it against the real application.

Jens · in Claude

Turn our biggest features into Holmes tests.

holmes · create_test

Via MCP3 tests

Book help at home

A family books a student for groceries and sees the booking.

Client portal

Chatbot answers a question

A family asks about their booking and gets a useful answer.

Client portal

Match a student

An admin matches a student to a new request.

Admin

Getting through the preview wall

The first real obstacle had nothing to do with testing. Staging and preview environments sit behind Vercel’s deployment protection, so Holmes was blocked before it reached the login page. A bypass secret in each product’s settings fixed it, and daily regression runs went live for the flows the business depends on most: the chatbot and booking.

Deployment protection

Holmes stopped before the login page

Bypass secret

Set in each product’s settings

Every night

ChatbotBooking

Testing work before it merges

At Maurice & Nora the real work happens on Vercel preview branches, one per pull request. Testing only after the merge to staging was too late.

“We’re big fans of the new feature testing, but the tricky part is that we’d have to merge before we can test.”

Jens Bogaert, translated from Dutch

So Holmes changed. It now runs when a pull request opens on a feat/holmes-* branch, reads the code change, tests what it touches on that PR’s preview, and posts a report back on GitHub. A failing test goes to Linear with its context attached, so a finding becomes a ticket without anyone retyping it.

Pull request

feat/holmes-* branch

Vercel preview

One per pull request

Holmes

Reads the change, tests what it touches

Holmes test report

Posted on the pull request

Failing test sent to Linear

Steps, screenshots and logs attached

Before merge

“I hooked Holmes up to Claude through MCP and turned our biggest features into Holmes tests. Worked really well.”

Jens BogaertFounding Engineer, Maurice & Nora

What it unlocked

Results the team trusts enough to act on

Running tests is only half the job; someone has to read the results. After a working session in August, the team asked for more logs on failing tests, failures at the top of a run, and a choice between running tests one at a time or in parallel.

That last one came from a real problem. With twenty tests in parallel against staging, Supabase auth locks and database timeouts made tests collide. The nightly run now goes one test at a time: slower, but every result means something.

20 in parallel

Supabase auth locks and database timeouts

One at a time

Slower, and every result is real

T

Tiemen in the shared Slack channel

“Thanks for following up and flagging it so fast! Jens has looked over his shoulder a few times because he thought you were watching behind him, that’s how fast you are.”

Translated from Dutch

Mobile, the next surface

From the start, Maurice & Nora said mobile testing was a must-have. They shared Android and iOS builds and gave Holmes access to their Expo account so it fetches new builds on its own. In September, Holmes ran its first smoke test on the client Android app.

New Expo build

Fetched automatically

Client app · Android

First smoke test, September

Smoke test

Open the app
Log in
Reach the home screen

“We’re really going to need to be able to do this, otherwise we’ll have to build something ourselves.”

Klaas, translated from Dutch

What’s next

Tests that live next to the code

Tiemen is already describing the next step: test definitions that live next to the code, so the coding agents that build Maurice & Nora’s product keep their tests up to date as well. That is where Holmes is headed with them: agentic testing that fits an agentic way of building software.

Know what’s broken before your customers do.

No credit card · First results in minutes