fleet-router — route your agents' model traffic through your tenant

Not generally available yet. The fleet-router is not open to tenants today: POST /library/router-token/<tenant> answers 503 router_not_yet_available and the router-dependent Data-Flywheel cannot be purchased. You can already pre-stage your BYO upstream with POST /library/router-config/<tenant>. The contract below describes the router once it is available.

The fleet-router is an Anthropic-compatible LLM gateway: it accepts POST /v1/messages (the only inference endpoint — there is no OpenAI-style inbound API) and forwards to a real upstream (Anthropic, OpenRouter, the Codex/Responses backend, or an OpenAI-compatible endpoint), choosing the model/subscription per a routing policy and attaching the upstream credential itself. Point an agent's ANTHROPIC_BASE_URL at it. Streaming (stream: true) works on every route (translated to Anthropic SSE for the OpenAI-family routes, passed through for Anthropic); once a stream has started there is no failover to another target.

There are two deployments of the same binary (dropdeadfu/fleet-router, fleet_router.py), and conflating them is the usual source of confusion:

only, uses the fleet's own subscriptions. Not something a tenant uses directly.

/library/router-token returns. The one *you* use. It is *intended* to require auth from every client and to serve per-tenant routing policy pulled from a4sx; read the status section before relying on either property. This spec is the a4sx-side contract.

Status and reachability — read this first

Two things that are NOT the same

The two auth layers

Identity: the token IS the tenant

Where the routing table lives — it's vault-backed

Turn it on (self-serve, in-harness)

Security posture

What the router answers with

Metering + pay-per-use billing

Managing the table from the app