Model Gateway

How it works

One connection point means one place to rotate a key, one place to set an allowlist, and one consistent record of what was spent.

Supported providers

Connect a provider

1. Sidebar → Models.
2. Connect provider.
3. Choose the provider. The form changes to ask for exactly what that provider needs.
4. Paste the credential, add a label (e.g. "Production OpenAI"), and save.

What Studio does next:

• Validates the credential against the provider.
• Encrypts it with AES-256-GCM and stores it in the internal vault.
• Registers the provider with the gateway.
• Marks the connection Connected.

Model IDs

Models are identified by a three-part ID:

openrouter/openai/gpt-4o-mini
│          │      └─ the model
│          └─────── the vendor who made it
└──────────────── the gateway provider it is routed through

This keeps things unambiguous when two providers offer the same model. Studio stores and sends the full ID unchanged.

Controlling which models your team can use

An admin can set an allowlist per provider. Only allowlisted models appear in the agent model picker.

Use it to:

• keep an expensive frontier model off limits for everyday agents,
• standardise on one or two models across the company,
• retire a model without hunting through every agent.

Per-member keys

Each member of the organization gets their own scoped gateway key, issued automatically when they join. It is used only for server-to-server calls and never reaches the browser.

The practical effect: model spend is attributable to the person who ran the agent, which is what makes the per-run cost in Audit traces and Usage meaningful.

Managing connections

Every one of these writes an audit row. See Audit logs.

Troubleshooting