How Claude Code fixes accessibility issues from your audit

AI coding assistants like Claude Code can fix accessibility issues directly in your code. But they need to know where the issues are. With the sitebrunch MCP server, Claude Code reads the findings from your audit, fixes the technical issues in your files and documents every fix back in sitebrunch. Content from the CMS and decisions about design stay with your team.

Key points

• First you test manually in the audit, then the AI fixes the findings through the MCP server, then you check again.

• The sitebrunch MCP server gives your AI assistant a CSS selector, a problem description and a solution suggestion for every finding.

• Technical issues work well, such as missing form labels, missing ARIA labels or a wrong heading hierarchy.

• The AI writes alternative text as long as it lives in the code. If it comes from a CMS the AI has no access to, the AI skips the finding and documents why. The same goes for color changes.

• The MCP server is included in all sitebrunch plans with audits, at no extra cost.

Why a person has to test first

An AI coding assistant sees your code. It does not see how people use your website. Whether the keyboard focus is visible, whether a screen reader reads content in a sensible order, or whether a dialog can be closed with the keyboard only becomes clear when you test the live page.

Automated testing tools only help so much. The UK government's accessibility team built a test page with 143 known barriers and ran common tools on it (GOV.UK). The best tool found 41 percent of the barriers.

That is why this division of work makes sense:

1. Find: A person tests the website manually, for example in a guided sitebrunch audit.

2. Fix: The AI applies the technical fixes in the code.

3. Check again: A person runs a re-audit to confirm the barriers are really gone.

What is an MCP server?

The Model Context Protocol (MCP) is an open standard developed by Anthropic. It connects AI assistants to external data and tools. Many AI tools support it by now.

The sitebrunch MCP server uses this standard to give your AI assistant access to your audit results. The AI processing runs in your own AI tool. sitebrunch only provides the audit data.

Which barriers AI can fix and which it cannot

Many of the most common errors on the web are technical errors in the code. The WebAIM Million Report 2026 automatically checked the home pages of the top one million websites. Some of the results:

• 51 percent had form fields without labels

• 46.3 percent had empty links

• 30.6 percent had empty buttons

• 13.5 percent had no document language set

The line is not between easy and hard errors, but along what the AI can access. Anything that lives in your code, an AI assistant can fix well once it knows where the issue is. Content that comes from a CMS is out of its reach. And design stays a decision for your team. In short: in the code yes, in the CMS no.

The reason for this limit: text from a CMS is not in your code files. The AI can neither see nor change it there, so your content team updates it in the CMS. Also review the alternative text the AI writes in the code. Text that describes an image incorrectly is worse for screen reader users than missing text. And a new color is a decision about your brand, not a technical correction.

A third limit concerns custom controls. Menus, dropdowns, sliders or tabs that only become controls through JavaScript need the right ARIA roles, states and working keyboard support. The AI can add attributes. But whether the element can really be used with a keyboard and screen reader afterwards only shows when you test the live page.

How the workflow works in four steps

Step 1: Run an audit. The MCP server works with the results of your sitebrunch audits. You test your website manually with the browser extension, step by step and with guidance. This is available from 49 euros per month, as a Kick-off audit or as a full audit with all 96 test steps against EN 301 549 and WCAG 2.2. Alternatively, our experts can run the audit for you.

Step 2: Connect the MCP server. In Claude Code, run this command in your terminal:

claude mcp add --transport http sitebrunch https://mcp.sitebrunch.com/

The first time you use it, your browser opens and you log in with your sitebrunch account. After that, the connection stays active. You can find the setup for Cursor, Windsurf, Codex, Claude Desktop and GitHub Copilot on the MCP server page.

Step 3: Let Claude Code work through the findings. Open Claude Code in your project and describe the task in plain language, for example:

"Get the critical findings from my latest sitebrunch audit for the home page. Fix the technical issues in the code. Skip anything that comes from the CMS or needs a design decision, and document why."

Claude Code then retrieves the findings from your audit by severity, gets the selector, screenshot and solution suggestion for each finding, looks for the right place in the code and applies the fix.

Step 4: Review the results. Every fix is added as a comment to the finding in sitebrunch, including the changed files. Skipped findings are marked with a reason, so your team can work on them directly. Review the code changes before you publish them. Then start a re-audit to check that the barriers are really fixed.

What to keep in mind

• Work in a separate branch. That way you can review every change before publishing and undo it if needed.

• Start with the critical findings. Let the AI fix the serious issues on your most important pages first, for example the checkout or the contact form.

• Test manually after the fix. Correct code does not mean the page is easy to use. Only a new test shows that.

• Test custom controls with a keyboard and screen reader. For these, correct code alone tells you the least.

• Plan time for the skipped findings. Alternative text from the CMS and contrast need someone from content or design.

FAQ

Which AI tools work with the sitebrunch MCP server?

All tools that support the Model Context Protocol over HTTP. These include Claude Code, Claude Desktop, Cursor, Windsurf, GitHub Copilot in VS Code and Codex. In GitHub Copilot, the MCP tools only work in agent mode.

Does the MCP server cost extra?

No. The MCP server is included in all sitebrunch plans that include audits. The cost of the AI itself depends on your own AI tool.

Can AI replace my audit?

No. The AI fixes issues that a person found first. Only a manual test shows whether a website is really accessible. But the AI takes a large part of the routine work off your team: finding the right place in the code, applying the fix and documenting it.