Build an agent that answers the phone

+ The pipeline is explicit, so interruption handling and turn-taking are yours to tune.
− A library, not a dashboard: no call logs or flow editor unless you build them.

+ WebRTC underneath means in-app and in-browser voice, not only phone calls.
− You run the media plane or pay for theirs - the framework alone is not a product.

+ Among the fastest real-time transcription APIs available, with diarization and keyword boosting built in.
− Developer-only; there is no end-user app, and accuracy on heavily accented speech trails Whisper-class models.

+ Latency low enough for a voice agent that can be interrupted.
− Fewer polished preset voices than ElevenLabs, and it is API-first.

+ The voice is the best part of the stack and it is included rather than bolted on.
− Minutes are priced as voice, not as orchestration - heavy call volume gets expensive.

+ Low-latency voice-to-voice NPC conversations with engine plugins.
− Usage costs scale quickly with players and the product has pivoted several times.

+ Books appointments and writes to a CRM without anyone writing code.
− The ceiling is the builder - unusual logic ends up fought rather than written.

+ Owning the whole pipeline keeps latency down and the per-minute price predictable at scale.
− A closed stack: the voices and models on offer are theirs, not yours.

+ The call-centre parts are already built: transfers, voicemail detection, recordings and transcripts.
− Less room to swap models than a toolkit; you work inside their flow builder.

+ Every layer is swappable, so a bad voice or a slow model is a config change, not a migration.
− You are renting orchestration and paying for the parts separately - the bill is four line items, not one.