MCP Server

import { Aside } from ‘@astrojs/starlight/components’;

@frontend-qa/mcp-server wraps the same scan/report pipeline the CLI uses and exposes it as five MCP tools, so an AI coding assistant can call them directly instead of parsing terminal output or reading JSON off disk itself.

`@frontend-qa/mcp-server` is currently `private: true` in this monorepo. It is **not** bundled into the `frontend-qa` CLI package (see [Design Decisions](/architecture/design-decisions/)), and its own standalone npm publish is not yet decided — see [Roadmap](/resources/roadmap/). Until then, run it from a built checkout as shown below.

Running it today

From a clone of the monorepo:

pnpm install
pnpm build
node packages/mcp-server/dist/bin.js

The server speaks MCP over stdio — register it with your client by pointing at that command, e.g. in an .mcp.json:

{
  "mcpServers": {
    "frontend-qa": {
      "command": "node",
      "args": ["packages/mcp-server/dist/bin.js"]
    }
  }
}

(This is exactly how this project’s own repository dogfoods it — see the root .mcp.json.)

Tools

Tool	Purpose
frontend.run_scan	Run a scan and return an issue-count summary
frontend.get_report	Return the full Report JSON for a run
frontend.get_issues	Return Issue[], filterable by severity/collector/category
frontend.get_console_errors	Return only browser.console issues
frontend.get_screenshot	Return a route’s screenshot as base64 PNG image content

Full argument schemas: MCP Tools Reference.

Why a separate server instead of shelling out to the CLI

The CLI and the MCP server both call into the same @frontend-qa/core and @frontend-qa/browser packages directly — the MCP server isn’t a wrapper that spawns the CLI as a subprocess and scrapes stdout. That means tool results are structured JSON (or an image content block for screenshots), not text an assistant has to re-parse.

Next

• MCP Tools Reference
• Architecture Overview