The mob, on tap — an MCP server for crowd-test.
Ask Claude Desktop, Claude Code, Cursor, or any MCP client to unleash a crowd of AI virtual users — impatient shoppers, confused seniors, keyboard-only users, chaos monkeys — on your website. They browse it in real Chromium, file findings, and hand back a damage report with a survival grade (S–F).
"Send the mob at https://staging.myapp.com and tell me what to fix first."
That's the whole workflow now.
pip install crowd-test-mcpAn LLM key is required in the server's environment: ANTHROPIC_API_KEY or
OPENAI_API_KEY.
claude mcp add crowd-test -e ANTHROPIC_API_KEY=sk-... -- crowd-test-mcp{
"mcpServers": {
"crowd-test": {
"command": "crowd-test-mcp",
"env": { "ANTHROPIC_API_KEY": "sk-..." }
}
}
}Related MCP server: stuntdouble-mcp
Tool | What it does |
| Send the crowd at a URL. Pick personas, add a random |
| The ten built-in ringleaders and what each one catches. |
| Preview the random mob a given |
| Fetch the newest full markdown report from disk. |
Findings can be cross-examined by up to three independent harnesses before they count against the grade — a skeptical detective agent, a raw Playwright probe, and a Microsoft Webwright agent. Automation artifacts get disputed instead of panicking you. Details in the crowd-test README.
Deeper layers need one-time extras:
pip install crowd-test[probe] && playwright install chromium # verify="cross"
pip install git+https://github.com/microsoft/Webwright # verify="tribunal"Runs take minutes, not seconds — every persona drives a real browser.
Start with 2–3 personas; escalate to mob=10 when you mean it.
Only test what you own. The mob is for your own staging and production sites, not other people's.
Reports default to ~/crowd-test-reports/<host>-<timestamp>/.
List the built-in ringleader personas and what each one catches.
Inputs: no named parameters. Returns: result
Preview the random mob members a given count/seed would generate.
Inputs: seed, count. Returns: result
Return the newest crowd-test markdown report under a directory.
Defaults to ~/crowd-test-reports, where run_crowd_test writes.
Inputs: report_dir. Returns: result
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 3 tools
Each tool targets a distinct action: listing personas, previewing mob members, and reading reports. There is no overlap in functionality.
All tool names follow the verb_noun pattern consistently (list_personas, preview_mob, read_report).
With only 3 tools, the server is thin but appropriate for a focused utility that augments an external test runner. The count is slightly under but reasonable for the scope.
The tools cover listing, previewing, and reading reports, but there is no tool to execute the crowd test itself, which is a notable gap given the server's purpose.
AI personas test your site like real users and return scored reports with fix prompts.
Browser-based QA for AI-built software. Test pages with real browsers via agents.
Deploy AI user personas to validate user journeys at scale. Find UX friction before real users do.
AI QA tester — real browsers scan sites for bugs, SEO, perf, and accessibility issues via chat.
