Chert open-sources a WebRTC framework to run agents on FaceTime
The YC startup's repo fields inbound calls, not just outbound bots — plus a demo where Claude pays a cent per page, and a cheaper 1M-token backbone.
Copy markdown
An agent that answers the call
Chert, a Y Combinator company, open-sourced a WebRTC framework on Sept 16 that lets an AI agent join FaceTime audio and video calls, with managed inbound-call handling built in. The hard part was never a bot that dials out — it's one that reliably picks up a live call, and that's exactly what the repo solves.
What you can build this weekend
Because it rides WebRTC — the same transport your browser already uses — you can point it at call screening, live tutoring, front-desk triage, or a see-what-I-see assistant a caller shows their broken router to. Clone github.com/cherthq/chert-facetime-opensource and wire in the model of your choice.
Elsewhere: agents that pay their own way
A ~300-line Hono demo uses the x402 protocol to charge AI crawlers a cent per page; on Sept 15 a Claude Code hook autonomously paid it in testnet USDC to read the content, capped at $0.05 per call and $0.25 a day. Flip HTTP's dormant 402 status on and your site can meter agents instead of blocking them.
Elsewhere: a cheaper backbone for realtime agents
DeepSeek V4.1-Flash (Sept 11) pairs a 1M-token context with off-peak output at $0.60 per million tokens and cache-hit input at $0.003, via KV-cache compression down to ~890 bytes per token. For an always-listening voice agent that keeps re-reading a long transcript, that cache math is the line between viable and not.