Feature request: additional tools

[michaelmckaycma]

Our Brightdata MCP-powered agent currently performs quite well, but often is unable to complete tasks, and I think it would help if it had some additional tools. In particular, I think the ability to execute arbitrary JavaScript (like playwright-mcp's browser_evaluate) could really help (the agent sometimes encounters shortcomings of ARIA snapshots, and is able to inspect the HTML instead but unable to interact with it). Something like browser_fill_form would also be useful - our agent takes some time/tokens to complete forms.

I can see that Brightdata MCP uses Playwright internally, and I wondered if it might even be possible to achieve parity with the playwright-mcp tools, which currently seem more advanced.

What do you think?

[meirk-brd]

Hi @michaelmckaycma,

Thank you for the suggestion - it sounds very useful!
At the moment, we support clicking and typing interactions, which effectively enables form filling. Do you feel that a more encapsulated fill_forms tool would provide additional value beyond what’s currently available?

By the way, the geolocation feature is nearly ready and will be released very soon. We really appreciate your ideas and your help in driving the evolution of Bright Data MCP!

[michaelmckaycma]

Hi @meirk-brd, to be honest the 'evaluate arbitrary Javascript' tool would be most useful - I noticed in some cases our agent (Claude Sonnet 4.5) 'asks' for it:

Since I don't know the exact structure from the HTML I examined earlier, let me try a direct JavaScript approach by executing JavaScript to click the button. However, since I don't have a direct JavaScript execution tool, let me try to find the button using the snapshot

Our agent seems to directly examine the HTML (using scraping_browser_get_html) quite frequently when the ARIA snapshot isn't clear, so to me this seems like a logical and powerful tool for an agent to have access to (it's available in playwright-mcp).

Our use for browser_fill_form essentially reflects the original use case described in this GitHub issue: microsoft/playwright-mcp#908. On certain websites (such as https://www.roboform.com/filling-test-all-fields, which the GitHub issue refers to as an example), our agent spends many minutes going through the form field by field. Moreover, we're using Strands so if we decided not to use browser_fill_form after all for some reason, it'd be easy for us to remove it from the agent.

What do you think?

[meirk-brd]

DONE — PR for your consideration: #96

I tested it on roboform.com:
https://www.youtube.com/watch?v=meWiwOHetr0

However:

when the ARIA snapshot isn’t clear

Have you tried fetching the ARIA snapshot with filtered set to false? That might be the issue. In both my implementation and Playwright’s, the form-filling logic relies on refs derived from the ARIA snapshot, so using it is required. In my testing, it works well with both the filtered and full snapshots.

Let me know if you have any additional ideas or suggestions for improvement.

[samtaylorcma]

Love this, thanks!