AI Provider Keys, Autonomy and Credits

Running coding agents against real machines raises three questions at once: whose model is being called, how much that agent is allowed to do without asking, and what the whole thing costs. AI Provider Keys, Autonomy and Credits is the VibeControls concept for answering all three in the interface rather than in a config file nobody reads. The three illustrations take them in turn — keys, gates, then metering.

The first illustration shows a keys panel listing six model providers one per row, each with a one-line note about which harness uses it and one or two credential fields. Several fields are masked, two of them marked as saved, while others sit empty, and the entries that need more than a secret also expose a base URL or a host field so an endpoint that speaks a compatible API, a routing service or a self-hosted model server can be pointed at directly. A subtitle states that the keys belong to a named agent rather than a central vault, and a footer note adds that sessions run through these harnesses are not billed against the platform's own credits — so a team that brings its own keys keeps its own bill.

The second illustration moves to an agent's security settings. One card governs remote task permissions with labeled toggles: running shell and script tasks from elsewhere is enabled, while remote file writes and deletes are denied, and a footer records who last changed the setting and when. A separate card offers session autonomy as three exclusive choices — plan only, with no edits or commands; accept edits, which applies file changes automatically but still asks before risky commands; and full auto, which carries a warning that it requires confirmation. An SSH configuration scan table beneath lists hosts with a risk column, giving the same page a posture check alongside the switches.

The third illustration puts cost where the work happens. An assistant panel docked beside the dashboard shows a credit meter with a progress bar and a remaining-of-total line for the month, a note that the balance is checked before every message, a short exchange about a degraded agent, and a line recording what the last turn charged and which model answered. The figures are sample data, but the placement is the point: the meter sits in the same panel as the conversation that spends it.

Read together, the three screens describe governance that scales with trust — a team can start agents in plan-only mode, then widen permissions and bring its own keys as confidence grows, with each step visible to whoever is accountable.