prufa-mcp — the QA agent for your vibe-coded app

Vibe-coded apps ship faster than anyone can review them. In June 2026 we audited 49 fresh Show HN launches — 38 had a critical bug on day one: a broken signup, a silent console error, analytics that never fired, a consent banner that did nothing.

Prufa is the agent that catches those before your users do. Point it at a URL and it audits the things humans skip when they're moving fast — broken flows, JS console errors, missing tracking, consent violations, security headers, mobile tap targets, accessibility — and hands back machine-verified findings, graded A–F. This repo is the open-source MCP server that wires that audit straight into your coding agent.

30-second demo

What an audit gives you

Ask your agent to audit https://yourapp.com and prufa_run_audit returns one JSON report. Findings are grouped into graded sections, each finding carries a severity, the impact (why it matters), and a fix hint. Real output, trimmed:

{
  "url": "https://yourapp.com",
  "headline": "2 warnings found",
  "counts": { "critical": 0, "warning": 2, "info": 5 },
  "sections": [
    { "label": "Works",     "grade": "C", "counts": { "warning": 2, "info": 1 } },
    { "label": "Fast",      "grade": "A" },
    { "label": "Found",     "grade": "A" },
    { "label": "Compliant", "grade": "A" }
  ],
  "check_results": [
    {
      "check_id": "ux",
      "findings": [{
        "severity": "warning",
        "title": "2 javascript console error(s) during page load",
        "impact": "Errors at load time often mean broken features visitors never report.",
        "evidence": { "count": 2, "sample": [
          "Access to XMLHttpRequest at 'https://api.fontshare.com/...' blocked by CORS policy",
          "Failed to load resource: net::ERR_FAILED"
        ]}
      }]
    },
    {
      "check_id": "mobile",
      "findings": [{
        "severity": "warning",
        "title": "13 tap target(s) smaller than 24px",
        "impact": "Fingers are not cursors — undersized buttons mean mis-taps on exactly the elements you want pressed.",
        "fix_hint": "Give interactive elements at least 24x24px of hit area (WCAG 2.5.8)."
      }]
    },
    {
      "check_id": "security",
      "findings": [{
        "severity": "info",
        "title": "no Content-Security-Policy header",
        "impact": "Without a CSP, one injected script owns the page — and every third-party tag you load is trusted completely.",
        "fix_hint": "Start with a report-only CSP and tighten from real violation reports."
      }]
    }
  ],
  "report_url": "/r/G82RpzTi_zn-o71_XoMLCprP7uvCQP87"
}

report_url is a shareable HTML version of the same report. The full payload also includes tracking, consent, seo/aeo, a11y, forms, and detected user flows — see the OSS surface below.

Install

The package is on PyPI. Install it globally with pipx (recommended — isolated venv, exposes the prufa-mcp binary on your PATH) or into a project venv with pip:

# Recommended — global install, isolated venv
pipx install prufa-mcp

# Or, into your project venv
pip install prufa-mcp

# Pin a specific version with ==, e.g. pipx install prufa-mcp==0.1.3

# Verify the binary is on PATH
which prufa-mcp
# Should print something like: /Users/you/.local/bin/prufa-mcp

You also need a free Prufa API key. The first audit is free, no card required.

1. Sign in at prufa.dev (Google OAuth)
2. Create an API key from the dashboard

Wire into your agent

The MCP server runs as a stdio subprocess, spawned by your agent on first use. The cleanest way to register it is claude mcp add (Claude Code's built-in command — it writes the config to ~/.claude.json correctly, which the ~/.claude/mcp.json path does NOT).

Claude Code (recommended path)

# Get the absolute path of the binary (use whatever `which prufa-mcp` returned)
PRUFA_BIN=$(which prufa-mcp)

# Add the MCP server. The token stays out of your shell history.
read -s -p "Prufa API token: " PRUFA_TOKEN && echo
claude mcp add \
  --scope user \
  --env "PRUFA_API_TOKEN=$PRUFA_TOKEN" \
  prufa \
  -- "$PRUFA_BIN"

Restart Claude Code (config is read at startup), then verify:

/mcp

You should see prufa listed as Connected, with prufa_run_audit and prufa_get_report as available tools.

Cursor / Cline / Continue (hand-edit .mcp.json)

In your project root or in ~/.config/Claude/ etc.:

{
  "mcpServers": {
    "prufa": {
      "command": "/Users/you/.local/bin/prufa-mcp",
      "env": {
        "PRUFA_API_TOKEN": "your-prufa-api-key"
      }
    }
  }
}

Restart the host app. The command path must be the absolute binary path (not ~, not $()) — those don't expand in MCP config.

Prefer config files to env vars? Drop your token in ~/.config/prufa/mcp.json instead — see ADVANCED.md.

Use it

In your agent:

> audit https://my-vibe-coded-app.com and show me the criticals
> run prufa on my staging deploy
> fetch the report for the audit I just ran

prufa_run_audit with wait=true (the default) blocks until the audit completes and returns the JSON report directly — typically 25–60s for a public page. If you set wait=false, the call returns immediately with the queued state plus a share_token you can poll with prufa_get_report.

What you get — the full agent surface

prufa-mcp exposes the whole product over MCP (44 tools). Point your agent at Prufa and it can audit pages, drive multi-step flows, watch for regressions, run chaos tests, run full-auto discovery, and manage the workspace + billing — no dashboard round-trip. Free/anonymous tools need no card; Pro tools return the API's 402 with a checkout link when you're not on a plan (the tool is visible, the paywall is server-side).

Audit & reports

Workspace, usage & conversion

Billing (returns a URL for the human to open — never takes a card)

Flows (describe a journey → reviewable spec → run)

Monitors (watch a URL or flow on a schedule)

Gremlin (chaos QA)