Skip to main content

Enable passthrough for Claude Code

Enterprise Tier

In passthrough mode a developer keeps their own Claude subscription (Max, Pro, Team, or Enterprise) and Anthropic bills them directly, while Tetrate Agent Router routes and observes the traffic. Three things have to be set up on the administrator side before that works.


Plan coverage: Max, Pro, Team, and Claude Enterprise subscriptions work, including Enterprise accounts that sign in with SSO.

Developers configure their own side separately, in Connect Claude Code.


What passthrough changes, and what it does not​

Passthrough changes how a request is authenticated. The developer's own Anthropic token is forwarded untouched and Agent Router adds no credential of its own.

It changes nothing about how a request is routed. The provider and the model are resolved exactly as they are for any other request, so the setup below is required whether or not the traffic is billed to the organization.

1. Enable an Anthropic provider, in the organization and on the project​

Passthrough works with Anthropic providers only, because it forwards an Anthropic token to an Anthropic API. There is no equivalent for Vertex, Bedrock, OpenAI, or a self-hosted endpoint. Those providers still route normally, using Agent Router's own credentials.

Enabling a provider for the organization does not make it usable by a project. A project has its own list of enabled providers, and the gateway serving that project only routes providers on that list. Enable Anthropic in both places: in the organization catalog, and on the project the developers use. Switch between them with the scope switcher at the top of the Admin Console sidebar.

Only one Anthropic provider is supported. A second one under a different name does not serve Claude Code, so an organization cannot run passthrough and managed Anthropic side by side: passthrough for developers with a subscription, and a managed key for agents and API callers, is not available today.

2. Set the provider's authentication type to API key​

Choose API Key and store an Anthropic key.

The authentication type does two things: it selects the credential used for managed traffic, and it tells the gateway which API format the provider speaks. Passthrough traffic uses the second: the developer's own token authenticates the request, and the gateway still needs to know it is talking to Anthropic.

The Upstream API field in the Add provider dialog sets which API the provider's upstream speaks. Left on Default for auth type, an API key provider is treated as OpenAI-compatible. Set the field explicitly for an Anthropic upstream configured under a name other than anthropic.

Since 0.4.0, moving a provider off the API key type later (to passthrough, or to retire the key) revokes the stored credential on the next configuration push. On earlier releases the stored key stayed live after the switch, so retiring it also means revoking it on the provider side.

3. Enable the models developers send, and assign them to the project​

A model that is not enabled, or not assigned to the developer's project, is rejected with 400 (model_not_found). This is the same check every request goes through; passthrough does not skip it. A 404 (model_not_routed) is a different case: the model is enabled and assigned, but the gateway has no route for it yet, which is normal in the minute after a model is enabled.

As with providers, enabling a model in the organization catalog makes it available to the organization, not to a project. Both steps are needed: enable it, then assign it to the project.

Claude Code has its own built-in list of models. It does not ask Agent Router what is available, so the models it offers and the models enabled in Agent Router are two separate lists. Enable the ones the developers use.

When a developer reports that passthrough is not working, check this first. They can only pick from Claude Code's list, so a model that is not enabled fails with nothing for them to change.

Checking it works​

Ask the developer for one request through Claude Code, then open Request Logs in the Console. A passthrough request is labeled passthrough, and shows the model that served it.

If the request never arrives, the problem is the developer's configuration. If it arrives and fails, work through the three items above in order.

If an enabled model is missing from /v1/models on the gateway, it is almost always the project step: either the model or its provider is enabled for the organization and not assigned to the project.

Passthrough and budgets​

Whether passthrough usage counts toward budgets is an organization setting, Count passthrough in budgets, in the Admin Console under System → Settings, on the Passthrough tab (from 0.12.0; a Super Admin changes it). With No, the default, passthrough is shown on each budget as Passthrough estimate (list price) and does not count toward the limit. With Yes, it counts toward every budget at list price, including a team's per-user budget. Until 0.13.0, a used-up budget refuses passthrough requests either way. From 0.13.0, a second setting, Passthrough can use managed models, decides whether that holds (see below). What counts, and how a per-developer budget behaves, is described in Budgets and Set budgets for users and teams.

Passthrough can use managed models​

From 0.13.0, a second organization setting, Passthrough can use managed models, controls what happens to a passthrough request for a model that has no passthrough route. With Yes, the default, the request is served on the organization's provider key, as before. With No, the request is refused with a 404 instead. The setting limits unplanned spend. It is not a security boundary, because a data plane older than 0.13.0 keeps serving such a request as before.

The setting cannot be changed in 0.13.0

In 0.13.0, the Admin Console's Passthrough tab under System → Settings shows only Count passthrough in budgets. The managed-models setting has no switch there, and a management key that calls the settings API gets 403. Every organization therefore runs with Yes, and a used-up budget refuses passthrough requests whatever Count passthrough in budgets is set to. The release notes for the release that adds the switch state when it can be changed.

The two settings together decide whether an exhausted budget affects passthrough requests. In 0.13.0 only the first two rows can occur:

Count passthrough in budgetsPassthrough can use managed modelsPassthrough request when a budget is used up
YesanyRefused or degraded by the budget action, because passthrough counts toward the limit.
anyYesRefused or degraded by the budget action, because passthrough can reach managed models.
NoNoNot blocked and not switched to a fallback model. Passthrough neither counts toward the budget nor reaches an organization-funded model.

The exemption applies only on data planes running 0.13.0 or later, and per project on data planes that serve one organization. A shared data plane keeps enforcing budget actions and reports Shared data plane as its enforcement state. A data plane that has not reported a version is not exempted.