multi-router.aidocs

Machine Identity

Calling Multi-RouterEnterprise

Token requirements, router selection, endpoints, limits, and how machine requests appear in Traffic.

A workload with a binding calls Multi-Router like any caller, with its platform token in place of an API key. The token goes in the Authorization header as a bearer token, and Multi-Router recognizes it as a signed token and checks it against the issuer and binding.

Token requirements

Multi-Router accepts a token that meets every one of these requirements:

RequirementDetail
audIncludes the issuer's token audience.
issEquals the registered Issuer URL.
subMatches a binding, which is enabled.
exp and iatPresent, with the token currently valid, allowing 60 seconds of clock difference.
LifetimeNo longer than the issuer's maximum token lifetime.
kidNames one of the issuer's signing keys.
Signing algorithmRS256, RS384, RS512, PS256, PS384, PS512, ES256 or ES384.

Endpoints

Machine identity works on every inference endpoint: Chat Completions, Messages and token counting, and the Claude Proxy, Codex Proxy and Gemini Proxy endpoints. With List available models, it also works on GET /v1/models. On a Claude Proxy with the Developer credential source and on a Codex Proxy without an organization key, the token goes in X-MR-Api-Token, since Authorization carries the provider login. Management endpoints, such as /v1/routers, are for API keys.

Choosing the router

A binding with one router sends every request to it. With several, each request names one, in the model field as multi-router/rt_9Hk3Ws7PzQ or in the x-mr-router header. A router outside the binding returns 404 the requested model or router was not found, the same as for a router that is absent. On the Claude Proxy's Bedrock and Vertex endpoints the message is Claude Proxy not found.

Rate limits

Each binding and subject has the same rate limit as an API key, 300 inference requests per minute, so each workload has its own allowance. See Rate Limit.

Traffic and budgets

Traffic records each machine request with its identity type, issuer, subject and binding, shown in the request detail as Subject type, Machine issuer, Machine subject and Machine binding. Machine issuer, Machine subject and Machine binding are also filters, and all four are optional columns. Machine requests count toward the budgets of the routers they use; a router for each workload gives each workload its own budget. See Request Detail and Budgets & Spend Control.

GitHub ActionsTroubleshooting
Was this helpful?