I run IronBase, a split fee recruiting marketplace and AI sourcing platform built entirely on Bubble. I come from recruiting, not software. Yesterday I connected the Bubble MCP to Claude with write access. Here’s what has shipped since then, in under 10 total hours of my time, most of it me previewing and approving rather than clicking around the editor.
A full candidate workspace on my sourcing tool. Saved candidates with a one click bookmark, personal folders with create, rename, delete, and an assign dropdown on every saved card, private notes on any candidate, and a Viewed badge with the date so you never waste a click on a profile you already opened. For agency accounts, the viewed activity rolls up across the team, so an owner can see what their recruiters have already covered without seeing it twice. Four new data types with privacy rules, a folder management popup, a saved panel with its own detail view, and all the workflows behind it. Manual estimate: 10 to 12 hours, easily my biggest build of the batch.
AI cost overhaul on my grading engine. Output trimmed on two graders, prompt caching restructured with a pre warm step and a second cached system block so 25 parallel grading calls actually hit the cache, deployed to Live, verified with real missions. My API bill dropped with zero loss in grading quality. Manual estimate: 8 to 10 hours, and honestly I would not have known where to start on the caching architecture.
Work mode filter on the job marketplace. The constraint had to land in 10 separate conditional data sources across tabs and filters. The MCP found every single one. Manual estimate: 3 to 4 hours of hunting, and I guarantee I would have missed one.
Job export feature. Select mode, select all, per card checkboxes, export action. Manual estimate: 4 to 5 hours.
New filter dropdowns on the marketplace. Styled to my design system, wired into the existing search constraints. Manual estimate: about 2 hours.
Search inputs on my submittals views. Two styled inputs filtering candidate cards by name, with the constraint correctly placed inside merged searches with ignore empty constraints. Built, previewed, working in about 20 minutes of my attention. Manual estimate: 2 to 3 hours.
A wiring audit I never would have done by hand. It read my element trees and workflows across two apps and caught that a pipeline status display described API calls that a dead page fired, not the live one. Finding that manually means clicking through every workflow step on every page. Manual estimate: 2 to 3 hours, realistically never.
That’s 30 to 40 hours of solo editor work done in under 10, starting the day I connected it, and the under 10 includes me asking questions and changing my mind.
What makes it work for me:
- Savepoint before every cluster of changes. The MCP creates them itself.
- One cluster at a time. It builds, I preview, I approve, then the next cluster. Never a giant batch I can’t untangle.
- It validates expressions before committing them, so I’m not debugging broken dynamic data after the fact.
- It reads the app before it writes. It pulled my existing search input pattern and mirrored it instead of inventing a new one.
If you’re a founder building on Bubble without a dev team, this is the biggest leverage I’ve found since Bubble itself.
