Skip to main content
Claude Code edits your code but cannot use a running app like a user, so it hands UI testing to the h/holo-desktop agent through holo mcp.

Claude Code ships a UI change, the desktop agent catches the regression, Claude Code fixes it and verifies again.

Nimbus Desk is a small support dashboard in the holo-desktop-cli repository. Its spec says that setting Status to Open on the Tickets page shows exactly #2042, #2040, and #2036. The app ships a plausible-looking bug: the filter compares each ticket’s status object to the selected string.
1

Start Nimbus

The app runs at http://localhost:5173. Credentials are on the login page: demo@nimbus.test / holo-qa-1.
2

Wire holo into Claude Code

With holo installed, sign in and install from another terminal, then open Claude Code from the Nimbus folder:
Claude Code loads the project’s .claude/skills/holo-qa/SKILL.md skill. It picks a spec from qa/, writes a self-contained task, and calls run_agent(task, agent="h/holo-desktop").
3

Ask for a change, plus QA

Claude Code makes the change, then runs the tickets-filter spec through run_agent as a black-box tester. The agent reports VERDICT: FAILED: with Open selected, the table is empty.
Nimbus Tickets page with Status set to Open and no visible ticket rows

The failing state the desktop agent reports: Status is Open, but no ticket rows.

4

Claude Code fixes the source

From that evidence, Claude Code finds the bug in examples/software_qa/src/pages/Tickets.jsx:
With the Priority dropdown added, the final code applies both filters. The QA spec must not change to make the test pass.
5

The desktop agent verifies

Claude Code reruns the same spec:
Nimbus Tickets page after filtering is fixed, showing three open tickets

After the fix: exactly the three open tickets.

​
Check a flow that works

The desktop agent also produces evidence that a working flow still works. Nimbus ships a passing spec, examples/software_qa/qa/chat-widget.md: from a signed-out session, the agent signs in, opens the chat widget, asks How do refunds work?, and checks that the reply mentions refunds processed within 5 business days. Ask Claude Code to run it without changing source. For CI or release checks: keep the Markdown spec stable, run the desktop agent against the built app, and store the final report plus screenshots as artifacts.

​
Troubleshooting

For setup and permission errors, see Troubleshoot.

​
Use the pattern in your project

Write a small Markdown QA spec for one user-visible behavior, wire holo into Claude Code, ask for a product change, and require Claude Code to verify the affected flow with run_agent: fix once, verify again. It fits authenticated flows, visual regressions, settings screens, local web apps, and native apps, anywhere source code alone is not enough evidence.