Model governance

Model access by plan

Oxagen routes all AI inference through the Vercel AI Gateway. No provider API keys are held in the Oxagen codebase or deployment — the gateway handles provider authentication.

Available models depend on your subscription tier:

All model calls are metered in credits at the per-token rate defined in the billing pricing table. See the full cost table for per-model credit rates.

Workspace model defaults

Workspace administrators can configure default model selection for their workspace via workspace.model.settings.write. This sets the default text model (and tier), image model, and video model applied to all agents running in that workspace, overriding individual user preferences.

To configure workspace model defaults:

1. Go to Workspace settings → Models.
2. Set the default text model tier and model name, image model, and video model as appropriate.
3. Save. Agents in this workspace will use these defaults unless a playbook explicitly overrides them.

Different workspaces can use different model defaults — for example, a cost-sensitive workspace can be pinned to Haiku while a high-capability workspace uses Sonnet.

BYOK (bring your own model API keys) is a planned feature and is not yet implemented. All inference currently routes through the Vercel AI Gateway using Oxagen-managed provider credentials.

Restricting model access per workspace

Workspace administrators can restrict which models are available to agents in their workspace, regardless of what the org subscription tier allows. This is useful for:

• Preventing high-cost models from being used in cost-sensitive workspaces.
• Restricting model access in workspaces that handle sensitive data to only pre-approved models.
• Enforcing consistent model behavior across a team.

Model restrictions are configured at Workspace settings → Models → Allowed models. Agents attempting to use a disallowed model receive a capability denied response, which is recorded in the audit log.

Model governance and org-enforced policies

Org administrators can set enforced policies on model access at Organization → Access → Policies. An enforced policy cannot be overridden at the workspace level. For example, an org can enforce that agent.code.execute only runs on Haiku (lower cost, lower capability) — workspaces cannot upgrade the model used for code execution without an org Admin changing the policy.

Model handle stability

Oxagen uses model handles (e.g., claude-sonnet-5) rather than hard-coded model slugs. Handles are resolved to the current gateway slug at request time via modelIdOf(). This insulates your agents from model naming changes in the underlying provider APIs.

If a gateway model slug is updated or renamed, handle resolution is updated centrally — your agents and playbooks continue to work without changes.