Exploring MCP Servers: Axe-MCP, cursor and Playwright for AI-Driven Accessibility Testing
I’ve been exploring a few open source MCP (Model Context Protocol) servers recently.
TL:DR
Here’s a youtube video with how I got on with the Axe-MCP server:-
The latest one that caught my eye was Axe MCP – an MCP compatible plugin for automated accessibility scanning. Shout out to Joe Colantonio’s Test Guild for his mega weekly series which brought my attention to this. Click the image to listen to the podcast. I definitely recommend connecting with TestGuild on LinkedIn and subscribing if you’re interested in the latest news.
I had a spare half an hour, so I thought I’d try it out.
The experiment
I had an existing Playwright framework which was basically the templated one you get when you install playwright, nothing fancy. I wanted to add a test to perform an accessibility scan using axe-core, driven by the MCP through my cursor IDE.
The Results
I was pleasantly surprised – in under 5 minutes and just two natural language prompts to my cursor agent this setup was able to:-
- Install and Add the MCP Server
- Add an accessibility scan test
- Execute the test
- Learn and iterate on the code – the initial test failed as chromium browser had not been installed, so this was automatically fixed (I was asked for permission)
- Execute the test again
- Summarise the accessibility findings in the accessibility report live in the chat
- Attach the stout accessibility report to the standard playwright test results html report
Take a look at the youtube video at the top of this page for the full details.
I think if you are interested in getting engineers to the point where they are adding value faster, provided the scans perform similarly to those generated with more traditional methods (be they “manual” or coded using Selenium commands for example) using MCP could be a way to get this off the ground a lot faster.
Not only that, but seemingly being partially self-healing could help reduce debugging time even more – that is, if you know what you’re doing and can course correct the Agent if it goes off track.
Things to be aware of
Its best to be very mindful of security of any open source MCP server. Security concerns are rife, and the importance of reviewing the code and keeping the human in the loop can’t be underestimated either.
Also, being created by a single user, this plugin is not officially affiliated with Axe from what I can see, which may cause maintenance and support issues down the road. I’m in awe of anyone who gives up their time to write open source software though, so huge kudos to Manosh Kumar for getting this over the line.
I haven’t experimented with this on other websites, so its possible I’m seeing a curated version of the output – I’d like to do a like for like with similar automated tests written in the traditional way to compare output if I were doing a full evaluation. UPDATE – I did try on Mark Winteringhams newly updated test website https://automationintesting.online/ and it successfully failed with critical accessibility issues detected:-
Finally, as with anything accessibility related, it isn’t possible to automate 100% of the testing – so please do not view this mcp extension as a replacement for traditional accessibility testing techniques.
Happy Testing!