Bubble MCP Feedback

Showcase
1

I tested the Bubble MCP with OpenAI Chat GPT / Codex (desktop app) with Sol 6.1 for about 1,5 hrs today on an existing Bubble app to see how far AI could help with understanding, improving, and maintaining an existing project:

  • Start/Setup: MCP Setup was easy as explained in the docs.
    Useful: When connecting the MCP, you can detail the access (e. g. limit to live / dev database, live / dev logs, etc.).

  • App analysis: In my first thread, I referenced the MCP just by saying “use the Bubble MCP” and it was loaded it into my conversation.
    It could very well explore the app’s structure (including visual data model) and explain how things fit together. That alone is useful when navigating an established project! Very good! By that, I could imagine to create a full documentation of an app just with the MCP.

  • Real changes directly in Bubble: improving page design and copy, creating editable elements, and making workflows easier to understand all through conversation. This was the most exciting part for me. I also added some screenshots to ChatGPT and asked them to implement it exactly like this in Bubble. Worked well! Tip: Open the test version in the built-in browser directly in Chat GPT to see your results appearing live. Also ask the MCP to validate its results via the test version (in the built in browser) if not automatically done. That was helpful as well.

  • Useful data exploration: I could ask questions about data and really do data analysis in natural sentences and get grouped results without manually building or exporting every query via the data tab. Some queries also cannot be build with today’s data tab since they referencing tables linked to another table. But that worked via MCP and Codex. Progress to the current data analysis options in the editor! Good!

  • Iteration: simple changes were straightforward. More subjective design work needed feedback. One interaction took three rounds before I was happy. Browser checks helped refine layouts across screen sizes. You can just annotate your findings in the built in Codex/Chat GPT browser and reference further changes from there.
    To make changes like you expect it, it is helpful to tell your AI tool “Remember to do this and that…” and every time it does a change via MCP, it actually considers it. Overall good!

  • Some limits remain (*Note: this is just from my 1,5 hrs test): creating or updating Bubble’s native App Text translations wasn’t supported by the available MCP. Workflow categorization also relied on naming conventions because folders weren’t exposed. I did not test workflow creation heavily. I did not do any analysis of my logs.
    My wishlist for data access would be to set a very granular access to every data type and field for both dev and live. Basically, just as for the AI agent currently built into the settings menu under “AI”. This setting would be super helpful for the MCP as well.

  • A few practical lessons: clear instructions, small batches of changes, and visual checks made the experience much better. I also set an explicit instruction for naming conventions for new elements etc.

Overall, a HUGE step for my personal Bubble setup in the AI age! I’m really excited about this update and the new options at hand and will definitely keep using it! Personally, I also like it more than the built-in agent but I will continue using both and gather experience. I also never used Buildprint and cannot make any comparison here.

2

this is a crazy good breakdown.

3

thanks! I already share this with the team. They are working on a lot of improvements for the MCP over the next couple of weeks :smiley:

4

Thank you, this is exactly the kind of thing we need to make the product better. Very much appreciated!

5

I’ve been trying out the Bubble MCP with Claude Code and really like it. One thing worries me, though.

When I connect, the MCP gets access to every app on my account, not just the one I’m working on. I have 16 apps and only want the AI touching one of them. The manual’s only advice is to “name the app you mean in your prompt,” which isn’t a real safeguard. Right now I’m relying on a check I set up on my end that blocks calls to the other apps. It works, but I’d much rather Bubble enforced it.

What I’d love is for each app to have its own MCP connection, set up from that app’s settings. That connection would only work on that app and couldn’t see anything else on my account.

As it stands, one wrong app name, whether it’s the AI’s mistake or a typo in my prompt, is all it takes for it to make changes to an app I never meant it to touch

6

Granular access control to apps, data types and even single data fields really would be a nice addition!