Stacker AI Builder: Three Weeks In — What It Actually Builds and Where It Still Needs You

TL;DR: Stacker's AI Builder has been public for three weeks. I've been using it to build actual portals and internal tools from spreadsheets and plain-English prompts. It excels at generating the boring-but-critical stuff (permissions, auth, data isolation) that vibe-coding tools leave to you. It stumbles on visual polish and complex multi-step logic. If you're building something customers or vendors need to log into, the governance advantage is real. If you want a gorgeous marketing site, look elsewhere.

What does Stacker's AI Builder actually promise?

The pitch is simple: upload a spreadsheet or connect Airtable, describe what you want in a sentence or two, and Stacker builds you a working portal or internal tool. Customer login, role-based permissions, data views, forms, the lot.

I first tried it the week it went public on July 1. I fed it a messy Google Sheet of client projects (columns for status, assigned contact, budget, deadline, documents) and typed: "Build me a client portal where each client logs in and sees only their projects, can upload files, and the internal team sees everything with edit access."

Forty seconds later I had a working app. Not a mockup. Not a component library. An actual portal with a login screen, two user roles, and my spreadsheet data sitting in properly linked tables.

That speed isn't unique anymore. The speed plus the fact that permissions and auth are already wired up is. That's the thing to watch.

What kinds of apps is it actually building well?

In three weeks of testing and watching community builds, a few clear patterns have emerged. The AI Builder does three things unusually well:

Client portals from spreadsheets. This is Stacker's home turf. If your spreadsheet has a column that maps to "which client owns this row," the AI detects it and builds the data isolation automatically. Client logs in, sees their stuff. Internal team sees everything. I built a vendor onboarding portal in under ten minutes — vendors upload compliance docs, the procurement team reviews them, nobody sees anyone else's submissions.

Internal CRMs and trackers. Sales pipelines, support ticket queues, project trackers. The AI understands these patterns and generates sensible layouts: list views with filters, detail pages with related records, status fields that actually change colour. I threw a recruitment tracker at it (candidates, stages, interviewers, notes) and it produced something usable on the first pass. Three or four tweaks in the visual editor and it was production-ready.

Simple approval workflows. Anything with a "submit, review, approve/reject" shape works surprisingly well. The AI sets up the status transitions and the permission logic so that submitters can't approve their own items. That's the kind of thing a human would forget to configure and the AI just does.

Where it's building these well, the common thread is structure. If your use case maps to a known pattern (portal, CRM, tracker, approval queue), the AI has seen thousands of examples and the output reflects it.

Where does it still need you?

Let me be blunt about the limitations, because they're real and they'll matter depending on what you're building.

Visual customisation is functional, not beautiful. The AI generates clean, workmanlike layouts. Everything is where it should be. Nothing looks broken. But if you're expecting the design fidelity of a hand-crafted Softr site or a polished Glide app, you'll be disappointed. The themes and branding controls help (upload your logo, set colours), but you can't radically redesign the UI without diving into CSS overrides. For internal tools and client portals, this barely matters. For customer-facing marketing experiences, it's a constraint.

Complex business logic needs manual wiring. If your workflow is "when status changes to X and budget exceeds Y, trigger approval from Z's manager," the AI won't generate that in one go. You'll need to build those automations yourself in the editor. The AI gets you 80% there — the data model, the views, the permissions — and then you do the last mile of logic. That's still a big time saving. It isn't magic.

Spreadsheet quality still dictates output quality. Feed it a clean, well-structured sheet with sensible column names and the result is solid. Feed it a mess of merged cells, inconsistent date formats, and columns called "Sheet1" and you'll spend more time cleaning data than building. The AI makes reasonable guesses, but it can't read your mind.

These aren't dealbreakers. They're exactly the trade-offs you'd expect from an AI that optimises for operational correctness over visual flair.

Why the governance story matters more than the AI

Here's the thing nobody writing about AI builders talks about enough: most of them generate code and walk away. The permissions, the authentication, the data access controls — that's your problem.

Stacker doesn't do that. The AI generates the app inside a managed environment where role-based access control is a first-class primitive. When you describe "clients see only their own records," the AI isn't writing comments in the code suggesting you configure this later. It's setting up the actual permission rules. They're live from the first build.

This matters because of something I've been tracking since March: the AI-generated code security problem isn't hypothetical anymore. Researchers have found 35 CVEs directly attributable to AI coding tools this year alone. The governance gap is real — 59% of organisations are running agentic AI but only 20% have governance in place.

Stacker's bet is that embedding permissions and access control into the AI generation process itself is the right answer. You don't add security later. It's part of the build. For anyone building tools that external users log into — clients, vendors, partners — that's not a nice-to-have. It's the difference between shipping and getting a call from your compliance team.

How does this compare to GlideOS and Bubble?

GlideOS entered open beta on July 14, two weeks after Stacker's AI Builder went public. It's a clean product. Upload a spreadsheet, get an app. The AI analyses your spreadsheet visually — it looks at formatting, macros, even takes a snapshot to understand layout context. That visual analysis is clever, and the output is prettier than Stacker's by default.

But GlideOS is positioning itself as an internal tools platform first. The permissions model exists but isn't the centre of gravity. You're building apps for your team, not portals for external stakeholders. Glide's pricing reflects that — external user seats get expensive fast.

Bubble's AI Agent is a different animal entirely. It's a conversational editor bolted onto the most powerful no-code platform on the market. You can chat with it: "Add a button here that triggers this workflow," "Change this colour to match the header." It understands pronouns and context. It's impressive. It's also still in beta with what Bubble themselves call "limited powers."

Bubble's approach is the mirror image of Stacker's. Bubble gives you infinite flexibility and the AI helps you build what you imagine. Stacker gives you opinionated, pre-baked patterns and the AI fills in your specific data and roles. One is a canvas. The other is a factory.

For the use cases Stacker targets (portals, CRMs, approval workflows with external users), the factory approach wins. You don't need infinite flexibility to build a vendor portal. You need it to work, be secure, and not require three weeks of configuring permission rules.

The takeaway

Three weeks is not a long time. But the pattern is already visible: Stacker's AI Builder is strongest where governance is non-negotiable and use cases are well-defined. If you're building something that external people log into and you don't want to think about auth, roles, or data isolation, it's the best option in the market right now.

If you want a beautiful public-facing app with custom interactions and polished animations, it's the wrong tool. That's fine. Tools that try to do everything do nothing well.

The bigger story is what this says about the AI builder market's splitting into two lanes. One lane generates code and leaves you to figure out production. The other generates governed apps with security baked in. The second lane is smaller, less sexy, and far more useful for anyone running actual business operations. Stacker is in that lane, and three weeks of usage suggests it's the right place to be.