A cloud voice AI running our own models on our own silicon. It hears how you say it, calls tools mid-sentence, and stops on a dime when you cut in.
One voice engine, three interfaces. Start with a plain request, move up to a reactive core that runs a real conversation.
Synchronous request and response. Text generation and function calling over the HTTP your stack already speaks.
REST docs →Bidirectional streaming, what the rest of the market calls "realtime." Token by token, barge-in, low-latency voice.
WebSocket docs →Monthly, through PayPal, cancel anytime. Full trace logs on every plan. Scale by concurrency, not by the minute.
Billed monthly through PayPal. We never see your card. Cancel anytime in PayPal; the current month is non-refundable. Compare with GPT Realtime and Gemini Live →
Not ready for a plan? Chip in once, get on the founders' wall, and keep your founding price for as long as the program runs.
Small models tuned for one job run on the phone's own AI silicon. When a turn needs more, it escalates to our cloud and drops back down. The system decides, turn by turn.
A focused model on the NPU answers before you finish asking. Works with zero bars. Nothing leaves the phone.
Reach, tools, or a harder question move the turn up the blade to the cloud. You never pick a tier by hand.
The full model family on our own accelerator fleet. Reasoning, tool calls, and speech on independent channels.
Public notes on on-device models, grounded decoding, and the mesh. Free to read.
Founding prices, new drops, and launch access. No spam.
HawkTalk AI, Seattle. Built and operated by Osprey. Our own models, our own weights, our own silicon. No reseller in the middle.