There are a few options out there, but the do the, “How do I rebuild my GPT on their platform?” thing, which just recreates the locked in aspect that the custom GPTs are exposing now. I’m building the migration infrastructure/system that answers, “How do I extract my intellectual property from Platform A, preserve it faithfully, and deploy it to Platform B, C, or my own infrastructure?” Copying a prompt into another product’s chatbot builder (even via an API or integrated connector) isn’t going to cut it without a lot of additional (beyond baseline) regression testing. It might not be traditional source code, but I agree and hear what your underlying concern might be … that your GPTs have activity that needs to be preserved (prompt engineering + documents + schemas + guardrails + workflows + APIs + test cases + tacit design decisions).

@Dr. John Elcik that is exactly what I’m building. 😊 Individual tenancies that are encrypted and yours, to then deploy your assets wherever. I’ve put out a data/help request to the group if you want to help 😊

@Sondra Verva Custom GPTs are still very much a thing. The UI has changed so that building/editing GPTs is currently web-only. You can use GPTs on mobile, but creation and editing require the web interface..

@Sondra Verva I have to apologise. 🙏🏻😊 I hadn’t actually created/edited my custom GPTs since June. 😞 You’re right though. The ability has been removed for all but Business, Enterprise, and Edu workspaces. Existing custom GPTs still function, however, which is why I didn’t notice. That was a sneaky change. 😡 I did some digging in the docs … no direct official announcement or release note, which is very poor show. Only thing vaguely suggesting a remova of/change to GPT’s was from an announcement on 22 April, “GPTs will remain available while teams test workspace agents with their workflows. Soon, we’ll make it easy to convert GPTs into workspace agents.” I dug a little more, and it looks like the change was being rolled out around 12 Aug when the help pages were updated to reflect the new policy (still no announcement though 🤷🏻‍♂️😡).

I ran over 12 different prompts of increasing complexity and have yet to find a failure in all 36 separate tests. 😳 Interestingly, I said to self, self, this will be easy … until it wasn’t. 🤣 I think that my flaw for the first 12 was human in nature as I used the same underlying structure for all: Named products > named sources > extract stated figures > follow explicit calculation rules. I created some, what I thought were complicated prompts, but GPT‑5.5 divided the work into clean retrieval and arithmetic steps. Increasing quantities merely made the same task longer. 😳 I’m then moved outside of standard query and on to an inverse identification problem: 1. Define a finite product universe, such as one manufacturer’s Australian range. 2. Never name the target product (leant my lesson on that after the first few tries). 3. Derive independent constraints from at least three different sources. 4. Include two or three extremely plausible near-matches. 5. Hide the decisive disqualifying fact in a manual footnote, compatibility table, regulatory record, archived specification, or regional SKU distinction (in other words, don’t make it obvious). 6. Ask only for the surviving exact model or article number. In theory, this made it materially harder because: - The model must search exhaustively rather than find the first plausible match. - The correct answer depends on intersecting constraints, including negative evidence. - Similar regional and superseded products contaminate search results. - Arithmetic cannot rescue incorrect product identification. - A wrong model number is unequivocally a wrong final answer. I then added a provenance chain (identify an unknown product from a replacement-part code, trace that code through a supersession notice, reject the apparent successor because of a regional electrical or mechanical incompatibility, and find the only currently purchasable compatible SKU). To remove common data structure with easily traceable many to many relationships, I then built and verified the entire finite candidate table, including every near-match and its precise reason for exclusion.

@Jason West wait, so you can’t simply create a project, upload the relevant reference files (stock and pice lists, google drive links for constantly fresh/live data, etc.), and then run voice within that project? 🤔 Or, is @Christina Ramirez referring to during the voice convo? I’m confused. 🤣