LLM Gateway

Every AI call in this fleet — every agent’s own reasoning, every text-to-speech call, and any other app that needs a model — goes straight to the outside provider it needs, using that app’s own scoped credential. There’s no shared proxy in the middle anymore.

Each consumer carries its own provider API key as a plain scoped credential — a compromised or leaked key only reaches whatever that one app was ever given a credential for. There’s no shared spending ceiling, shared rate limit, or centralized spend report anymore; see AI Provider Access & Cost Control for what that tradeoff actually cost and bought back.

The orchestrator agent still needs a way to reach the weather/time/currency agents without hard-coding three separate addresses, so those three sit behind one small internal router — see Agents Router for its actual deployment shape. This is deliberately a much narrower thing than the old gateway was: it does no reasoning, holds no model registry, and enforces no budget — it’s pure address-bundling, trusted only because network policy restricts who can reach it, not by any credential on the call itself.

The old gateway used to stamp every call with a shared identifier and forward it to the fleet’s tracing system, so a whole podcast episode’s script-writing call and every narration call it triggered would show up together as one trace. That forwarding step no longer exists — see Langfuse for what that means today.

See Agents Router for the delegation router’s actual deployment shape, and AI Provider Access & Cost Control for the full reasoning behind giving up the shared gateway.