Agents are the users now

Most hosting is built for a person sitting at a dashboard. The person clicks, reads a chart, copies a connection string into an app, and goes to lunch.

That person is increasingly not there. A coding agent wrote the app, and the app needs a database, and the agent is the one asking for it at 2am while you sleep.

So we built Kraftify for the agent first. Not as a second interface bolted onto a human product. As the primary user, with the dashboard as the convenient side effect.

This is what that decision looks like in practice.

We believe that, in 2025, we may see the first AI agents "join the workforce" and materially change the output of companies.

Sam Altman, CEO of OpenAI

The same calls, whatever holds them

There is one API underneath everything. The CLI calls it. The dashboard calls it. And the MCP server exposes it as ten tools your agent can call directly, six of which only read.

Nothing in the right column is a wrapper around the left one. They are the same operations, which means a script you write today and an agent run next month cannot disagree about what happened.

The pricing deserves its own mention. Rates live at a stable URL in plain markdown, because the reader is as likely to be a model parsing context as a human comparing plans. A price an agent cannot read is a price the agent will guess at, and guessing about money is exactly what you don't want it doing.

Trust through bounds, not through watching

Being pro agent does not mean giving the agent everything. It means giving it exactly enough, then bounding the rest so you can stop watching.

The ceiling is the load bearing piece. A create that would cross it is refused, and nothing the agent holds can raise it. Not a flag, not a tool, not a clever prompt. The number moves only where a human moves it.

Around that, the defaults do the rest. Start the agent read-only and judge what it can see before it can do anything. Scope its token to the tools the job needs, and never to the one that mints tokens. There is no console write to hand over, which is not a missing feature. It is the shape of the product: the agent acts through the same narrow calls you would make yourself.

This is also where the last post connects. An agent acts as you, under your ceiling, within your scopes. Judging behavior instead of demanding identity turns out to work for software exactly as well as it works for people.

Agents are not seats

There is no per-seat pricing here, because the concept makes no sense anymore. Is your agent one seat? Are its fifty subagents fifty seats? Is the overnight run that deploys while you sleep a user?

The unit is the target: one box, one flat monthly rate, whether a human clicked it into existence or an agent created it at 2am. Pricing that counts humans breaks the moment the humans stop being the ones doing the work.

So to put it plainly: we are pro AI in the specific, checkable sense. Ten tools, six read-only. One ceiling nobody can talk their way past. Docs and prices in a form a model can actually read. No seats, no surcharges for acting through an agent, no human checkpoint inserted where a bound would do.

Build the bounds. Then let go.