Thanks — I’m sorry you hit this on startup. I’m responding to your question about whether the Build plan is subject to a WebSocket connection rate / concurrency limit on the inference gateway, and whether you can increase it or disable prewarming to avoid simultaneous connections. Short answers • The Build plan enforces LiveKit Inference concurrency limits (Build provides 5 concurrent inference connections). Simultaneous STT + TTS WebSocket opens during startup can exceed concurrency or burst limits and return HTTP 429. • There isn’t a documented SDK option to disable the SDK’s prewarm/connection behavior or to change the gateway’s per-second connection throttling. • You can request higher inference concurrency limits via the LiveKit dashboard to reduce the chance of 429s from burst connection attempts. What you can do next • Temporarily avoid simultaneous model connections by creating the AgentSession with only the models you immediately need, then lazy-init the other inference clients after startup (e.g., create/connect to TTS after STT is established). This staggers WebSocket opens and often avoids 429s. • Request an increase to your inference concurrency from the dashboard so the gateway allows more concurrent model connections for your account. • If you can share the 429 response body (the 495-byte JSON) I can help interpret whether it indicates a concurrency limit or another throttle reason. Also tell me whether this happens reproducibly on first startup or only after other activity in your account. If you’d like, I can walk through a minimal code pattern to lazily initialize STT/TT S connections in your AgentSession so they don’t open simultaneously. Which approach do you prefer? Sources: https://livekit.com/pricing.md| | Pricing | LiveKit | Build voice, video, and physical AI | LiveKit | Pricing | LiveKit | Server options | LiveKit Documentation, with interactive elements

Thanks — I see the exact 429 WSS handshake error for the STT endpoint you pasted. I’m sorry this is blocking startup. What’s happening • The gateway is returning HTTP 429 at the WebSocket handshake for

/v1/stt

(Content-Length: 495 JSON body). This indicates the gateway is rejecting the connection due to throttling/concurrency limits rather than a protocol bug in your client. • On the Build plan LiveKit enforces inference concurrency limits (Build provides 5 concurrent inference connections). Opening STT and TTS WebSockets simultaneously during AgentSession startup can count as concurrent/burst connections and trigger 429s. Immediate mitigations you can try • Stagger the connections so STT and TTS don’t open at the same time (lazy-init one after the other). This avoids the burst that triggers the 429. I can provide a short code snippet to do this in your agent if you want. • Create the AgentSession with only the inference clients you need immediately, then connect the others after startup succeeds. Longer-term options • Request a higher inference concurrency limit from the dashboard for your account so the gateway allows more concurrent model connections and reduces 429s from burst opens. Information I can use to help further • Paste the 495-byte JSON response body (it likely explains the throttle reason). • Confirm whether this 429 happens every first-run attempt even on a fresh account, and whether it reproduces if you only enable one of STT/TTS. Which would you like next — a brief code pattern to lazily initialize TTS after STT connects, or help interpreting the 429 JSON body? , with interactive elements