# Limits & Quotas Every timeout and its default, the context cap, audio-format requirements, rate limits, and the browser/device support matrix — in one place. The limits that affect a running session, collected so you don't have to hunt for them across pages. ## Timeouts | Timeout | Default | What happens | | --- | --- | --- | | Idle session timeout | ~2 min | A session with no interaction ends automatically. Keep it active or reconnect to persist. | | Error lifetime | Until next start() | The error state is sticky — it never clears on a timer, so a dead-end screen rendered off error.fatal stays put until you reconnect. | | Session token expiry | ~8 s | Tokens are single-use and very short-lived; the SDK requests a fresh one via getSessionToken on each connect. | | Connection setup | 5–7 s first / 2–3 s cached | First connect downloads and initializes the Rive avatar; later connects load it from cache. | > **Runaway usage:** metering accrues while the **avatar is speaking**, so a loop > that makes the avatar talk repeatedly burns usage fast — guard against it. > Separately, every open session counts against your concurrency limit (5 open > connections), so end sessions you no longer need. ## Input caps Exceeding any of these emits a **non-fatal** `INVALID_INPUT`. The call is dropped; the session is untouched. | Input | Cap | Notes | | --- | --- | --- | | addLiveContext() context | 2,000 chars | Per call. Summarize or chunk before sending. | | speakText() text | 5,000 chars | Per call, in tts mode. | | speakAudio() audio | 15 MB | Per call, in audio mode. | | Greeting message | 400 chars | Clamped to match the server, so a longer message is truncated rather than rejected. | ## Built-in feature limits Defaults inside the SDK, most of them adjustable in config. | Limit | Value | What it governs | | --- | --- | --- | | Transcript entries retained | 50 | Configurable via transcript.maxEntries. Older lines fall off the front. | | Connect tone ceiling | 15 s | Hard cap regardless of loop, so a wedged connect cannot leave the ringback droning. | | Greeting lead-in | 500 ms | The beat after the session goes live before the avatar speaks, so the ringback's fade can finish. | | Push-to-talk pre-roll | 200 ms | Already-captured audio released on press — people start talking just before their thumb lands. | | Push-to-talk release tail | 250 ms | Real audio still sent after release, so the last syllable is not cut. | | Push-to-talk max hold | 2 min | A hold this long is a stuck key, not a person. The turn is closed. | | Camera frame buffer | 12 s | Rolling window, in memory only. Older frames are discarded on every capture. | | Camera frames per look | 3 | Newest last, at 3 fps — roughly the last second. | ## Audio format (audio mode) `audio` mode lip-syncs to audio you supply from another provider. Send **PCM** audio in the exact format below; other formats or sample rates won't drive the animation correctly. See [Modes](/docs/sdk/react/modes) for the per-mode contract. | Requirement | Value | | --- | --- | | Format | 16-bit signed PCM, little-endian (PCM_S16LE) | | Sample rate | 16000 Hz | | Channels | Mono (1) | ## Rate limits | Surface | Limit | | --- | --- | | Session token endpoint | 120 requests / min per API key | | Concurrent sessions | 5 open connections (per IP) | All plans currently allow **5 concurrent sessions** (5 open WebSocket connections). Manage your account in the [Platform](/docs/platform/avatars). ## Browser & device support | Platform | Status | | --- | --- | | Chrome / Edge (desktop) | Supported | | Safari (desktop) | Supported — audio must be unlocked by a user gesture | | Firefox (desktop) | Not yet tested | | iOS Safari | Supported — autoplay requires a user gesture; see FAQ | | Android Chrome | Supported | | React Native (Expo) | Supported via a dev build (not Expo Go) | > Building for mobile? The React Native SDK has its own permissions and echo > guidance — start at the [React Native Overview](/docs/sdk/react-native/overview).