Bring your own model: why your AI key should live on your device, not on ours

You already know the shape: most workplace AI platforms sit between your people and the model. Requests route through the vendor’s servers, credentials live in the vendor’s vault, and inference comes back on the vendor’s bill — one box in the middle of the architecture diagram, familiar enough that nobody questions it anymore.

This article questions it. The box has good stated reasons to exist — convenience, “enterprise controls,” a single invoice — and one unstated reason that explains more than the other three combined: whoever owns that box owns the relationship, the data flow, and the margin. What follows is what the box quietly costs you, and the fact that nothing technical requires it to exist at all.

The old way: the vendor in the middle

The standard architecture of workplace AI is a proxy. You sign up; the platform provisions model access on its keys or stores yours server-side; every request from every employee routes through the vendor’s infrastructure; and inference comes back to you metered, marked up, and bundled into the seat price. Three consequences follow, and they are structural, not accidental:

• Your data flow is their asset. Everything anyone sends to a model transits systems you don’t control, under terms that can change, at a counterparty that can be breached, subpoenaed, or acquired.
• Your model choice is their roadmap. You use the models the platform integrated, when it integrated them, at the prices it set. The frontier moves monthly; your procurement cycle doesn’t.
• Your exit is their leverage. The proxy accumulates configuration, history, and habit. Leaving means re-plumbing everything — which is precisely the point.

The alternative: bring your own model

None of that middleware is technically necessary. Model providers issue keys directly; the missing piece was never access — it was a workspace willing not to insert itself into the connection. That is the design choice CloudTop — the AI workspace you own — makes, and it is worth stating in its exact, unshortened form: CloudTop never holds your model key, never proxies the traffic, and never bills for inference. The key sits in your own settings, on your own device, and only what you ask the AI to look at leaves — to the model you picked.

What per-task model choice actually buys you

Once the connection is yours, model choice stops being a procurement decision and becomes a routing decision — made where the work is, at the moment it matters.

And “you” means each person, not a committee. In any real company the preferences already differ: one colleague is a heavy Claude user, another gets better results from ChatGPT or Gemini, a third wants a small, specialized model for one narrow job — different habits, different strengths, different tasks. The usual platform answer is to standardize on a single model and call everyone else a training problem. In CloudTop there is nothing to standardize: each person connects the model they actually prefer, switches per task, and works in the same workspace, on the same shared data, all the while.

This is also the honest answer to the frontier problem. Models leapfrog each other every quarter. Under the middleman model, each leap is a vendor negotiation. Under bring-your-own-model, it is a Tuesday: paste the new key, point the tasks that benefit at it, keep everything else where it was. Your workflows, your data, and — with shared workspace memory — your team’s working knowledge all stay put, because none of them were ever stored inside a model vendor’s account.

The enterprise objection: “we need one throat to choke”

The most serious pushback to bring-your-own-model comes from finance and IT, and it deserves a straight answer: doesn’t the middleman at least consolidate billing, provisioning, and spend control? Consolidation, yes — control, no, and the two keep getting sold as a bundle. Model providers already offer organization-level keys with member management, usage dashboards, and spend limits; “one invoice” is available from the source, without inserting a reseller into your data path to get it. Provisioning collapses into workspace policy: the organization’s key lives in managed settings, folders route work to approved models, and an employee’s first day requires no per-vendor seat ceremony. And spend visibility actually improves, because you see the provider’s raw metering instead of a bundled seat price engineered to make the margin unfindable. What the middleman uniquely offered was never the invoice — it was the comfort of a single vendor to blame. Weigh that comfort honestly against what it costs: your keys in their vault, your traffic in their logs, and your exit in their hands.

“But the middleman gave us control”

The usual defense of the proxy is governance: if everything routes through one box, at least someone can watch the box. It is a real need answered in the wrong layer. Watching traffic tells you what already left; governing the workspace decides what can leave at all. In CloudTop, control lives where the data lives — which folders a model may read, which applications expose which functions, what requires a human click — and every AI action carries a visible tool trace. You get the oversight the proxy promised, without paying for it in keys, margin, and lock-in. The deeper mechanics are in credentials never leave the device.