Bring your own key: what it costs and what it protects

"Bring your own key" means the model calls run on the API keys you already pay for, not on ours. It is the governing principle of Cortex66: your own AI first. This article says plainly what that costs, what it protects, and what the system does with a key once it has one.

What it costs

• Model calls are billed by your provider to your key. Cortex66 never adds a markup on model calls.
• The cap is shown before each run. Every task starts with a receipt that names the most it can cost. Nothing runs until you accept it, and the cap is reserved before any paid call leaves.
• Connecting a key runs one capped test. The test proves the key works and reads the provider's own signals; it is small and it is capped like everything else.

What you pay is therefore what your provider charges for the tokens your tasks used, inside caps you set. The cost cap page shows how the cap is reserved and what happens when it is reached.

Which keys work

OpenAI, Anthropic, Gemini and xAI keys, and local models running on your own machine, can sit side by side in one workspace. You do not need an account with every provider: add only the keys you have. A task that needs a model you did not add is shown as blocked, with the reason.

Private work can stay on a local model while the rest uses a cloud model. Each receipt says which one ran and why.

What it protects

A key is a secret, and Cortex66 treats it as one:

• It is stored as a secret and used only for the tasks you accept.
• It is never shown back to you or to other members of the workspace; only its last four characters are ever displayed, so you can tell keys apart.
• It is never written into a log, a receipt or an event; the receipt names the model and the lane, not the key.
• Pausing or revoking a connection stops its use at once; nothing already running is sent again.

A second fence on the provider's side

Our ledger is the authority on what a run may spend. For providers that support project spend limits, the connection screen recommends setting one there too. It catches what a ledger cannot: a bug, or a key that leaked somewhere else. Two fences, and the inner one never relies on the outer.

Choosing what a provider may read

Not every provider should see every task. Cortex66 lets you say, per provider, what it may read, and keeps a task away from a model you did not trust with that data. The receipt shows the choice, and a task that cannot run under your settings is blocked with the reason instead of running somewhere you did not intend.

What a key never does

• It never runs a task you did not accept.
• It never pays for another member's work outside the caps you set.
• It never leaves the workspace in an export, a brief to another AI or a page.

Read next: which model for which task and one cap for every AI.