Early on, my research revealed Gemini support for Scribld might not be possible. For months I put off cross-platform support, because I took it as fact.
And my research was technically right. Gemini and ChatGPT can't do the exact same thing as Claude. Our Claude setup is an instruction file plus the model fetching URLs with an auth header, and ChatGPT's browsing can't send custom headers. Gemini's consumer app can't call outside APIs at all.
But "it doesn't work" and "it doesn't work that way" are different statements, and I'd been treating them as the same. The goal was never "make ChatGPT fetch a URL." The goal was: you say "scribble this down" in whatever tool you're working in, and the note lands on the right board in your workspace.
MCP does that. One remote server, three small tools, and every client that speaks the protocol can save and search your notes. Claude, Claude Code, ChatGPT, Codex, Gemini CLI, Cursor... MCP covers all fronts, plus what comes next. We didn't build integrations, we built the UI for the memory layer, our goal from day one.
The plan said five weekends
When I spec'd this out with my AI tool I estimated ten or fourteen nights of work.
It shipped in a day. And I want to be honest about how, because it's the whole reason Scribld exists: I build with AI agents. I spun up an agentic team to run the refactor, the server, the install flows, and the verification passes while I acted as the human in the loop.
That's not a flex. It's the thing I keep learning about building in 2026: the constraint isn't hands anymore, it's decisions and time. Which makes it exactly the kind of week where your decisions pile up faster than you can remember them. So yes, I scribbled the whole build down as I went. The product got built while eating its own dog food.
Rug-pull number one
The night we wired up ChatGPT, I opened the connector dialog and the option the whole plan depended on was gone. When I researched this in early July, ChatGPT's Developer Mode accepted a pasted API key. Somewhere between then and the moment I clicked Create, OpenAI removed it. OAuth or nothing.
So we built OAuth that night. A small one: a consent screen on our own login, short-lived codes, and one design decision I'm glad we made under pressure. The "token" ChatGPT walks away with is your existing Scribld key, delivered a fancier way. Revoke your key in Settings and every platform loses access at once, ChatGPT included. One kill switch, no second credential system to babysit.
Rug-pull number two
Google got there before we did. Two weeks before the build, Gemini CLI stopped accepting Google-account sign-in for individuals entirely. The CLI still works with an API key, the free tier covers it, so that's the documented path now.
The funny part: my own Google Workspace refused to issue me a working key. Every admin toggle on, brand-new project, denied anyway. "Contact support." I found another way to verify Gemini support.
What I took from it
Two platforms changed their auth rules between my first line of code and shipping my MVP. That's the pace of AI right now, everything moves and nothing sits still for long. The concept survived because it leans on MCP, an actual standard, instead of a platform-specific trick. A standard is the closest thing to solid ground in a space moving this fast.
And the "might not be possible" thing... the research was right, and it didn't matter. Worth remembering, next time the research says no: check which question it's actually answering.
Anyways. If you build across models, your project memory doesn't have to reset at each one anymore. Same trigger phrase, same boards, same key, whether you're in Claude, ChatGPT, Codex, Gemini CLI, or Cursor. Setup takes about two minutes per tool: scribld.io/setup.
Scribld is the persistent memory layer for building with AI. Say “scribble this down” and the decision gets filed, searchable, across every session. Try it free for 14 days, no card required.