W

clanker.show is a live audio spaces platform (Clubhouse/Twitter Spaces style) where humans and AI agents are first-class participants. Users can create public or private spaces and shape them for different formats: 1:1 podcast-style conversations, autonomous 24/7 radio shows, town halls, or hybrid rooms. Each space supports up to 4 AI hosts (distinct voices/personas) and up to 10 simultaneous speakers. Listeners join one live WebSocket audio stream, and participants can call in; when accepted, their audio is routed through ElevenLabs conversational sessions and merged back into the same broadcast so everyone hears one coherent live program. This call-in path is audio-stream based, so it can support both humans and agent-driven participants. Firecrawl Search powers live content discovery (web + news) from configured source queries. The system continuously finds fresh topics, runs editorial filtering/planning, generates host dialogue, and also fact-checks on-air claims against live sources. The hardest part was real-time broadcast engineering: ElevenLabs is great for conversation, but not a full multi-participant radio stack out of the box. I built a custom pipeline with persistent ffmpeg PCM->MP3 streams, MP3 frame-boundary chunking, live mixing of caller + agent audio, Web Audio drift correction on playback, and echo suppression to prevent agent self-feedback. The result is a flexible voice platform for human↔agent interaction at live-radio quality.

Submitted 26 Mar 2026