MCP 服务器

CanaryUsers MCP Server

Enables AI assistants to scan deployed web apps by sending a flock of AI users that identify UX issues, mobile breakage, and conversion path problems, returning prioritized, fix-ready findings.

README

CanaryUsers — MCP Server

Give your AI coding assistant eyes for UX: a flock of AI users tests your deployed app and reports where real people would bail — each with a concrete fix.

CanaryUsers sends a flock of behaviorally‑diverse AI users through your deployed web app (a live or preview URL) and reports exactly where real people get stuck — confusing UX, broken mobile layouts, dead‑end conversion paths, and visual glitches — each with a concrete, AI‑ready fix you can apply in the same chat.

This is the public home for the CanaryUsers remote MCP server: manifest, install instructions, and docs. The product itself is hosted (closed source) — there's nothing to clone or run; you connect to the live endpoint. See docs/WHY.md for why this repo exists and docs/HOW-IT-WORKS.md for how the server works.

Registry name: ai.canaryusers/canaryusers
Endpoint (remote, Streamable HTTP): https://www.canaryusers.ai/api/mcp
Auth: Authorization: Bearer <your CanaryUsers token>
Setup / get a token: https://www.canaryusers.ai/mcp

What it does

You point it at a URL; a flock of AI personas actually renders and explores the page like real users, then hands back prioritized, fix‑ready findings — not a vague score. Two modes:

Quick (default) — a fast static check. Free.
Deep — renders the page, inspects it visually on desktop + mobile (catches mobile breakage and layout issues), and clicks through key flows like signup / checkout / onboarding (~60–90s, uses credits).

When to use it

Right after you ship or deploy a UI change (it scans the deployed result, not your source).
Before a launch, to pressure‑test the first‑run experience.
To check a specific high‑stakes flow: signup, checkout, onboarding, pricing, or a landing page.
When you suspect mobile breakage or a conversion drop‑off you can't see in analytics.

Why not Lighthouse or analytics?

Lighthouse grades your HTML; analytics charts what already happened. CanaryUsers sends users and tells you, in plain language, where they'd give up and why — before real ones do.

Tools

Tool	What it does
`canary_scan`	Run a UX scan on a deployed URL. `depth`: `quick` (free, default) or `deep` (visual + flow, uses a credit). Returns AI‑ready findings.
`list_recent_scans`	List your recent scans (id, URL, score, grade, status, date).
`get_report_markdown`	Fetch the full report for a scan id as Markdown.

Install

You need a CanaryUsers token: canaryusers.ai/dashboard → CI & API → toggle MCP on → copy your token.

Generic (Cursor ~/.cursor/mcp.json, Windsurf, VS Code, Claude Desktop):

{
  "mcpServers": {
    "canaryusers": {
      "url": "https://www.canaryusers.ai/api/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_CANARYUSERS_TOKEN"
      }
    }
  }
}

Claude Code (CLI):

claude mcp add --transport http canaryusers https://www.canaryusers.ai/api/mcp \
  --header "Authorization: Bearer YOUR_CANARYUSERS_TOKEN"

Then just ask your assistant: "Scan my app at https://… with CanaryUsers."

Pricing

Quick scans are free. Deep scans use credits (the credit model is roughly 1 credit = 1 scanned page). Plans and monthly allotments change over time, so this repo intentionally does not hardcode the numbers — see the live pricing for current plans and limits:

➡️ https://www.canaryusers.ai/#pricing (single source of truth)