Bolna API + MCP vs CueLight

Bolna API + MCP

Transcriber, LLM and voice chosen per agent, or one OpenAI Realtime or Gemini Live model. Unsigned webhooks and tool calls, IP allowlisting only.

CueLight

CueLight gives a voice pipeline a local end-of-turn probability every 20 ms from audio alone, with a 1.9 MB model and Apache-2.0 weights. It is one day old, English only, installs from Git rather than PyPI, has no CI or changelog, and its accuracy figures are the author's own.

Bolna API + MCP

1. Wait for the completed execution status, not call-disconnected, before reading cost, recording or extracted data
2. Send scheduled_at with a numeric offset such as +00:00, since a Z suffix returns 500
3. Back off exponentially on 429, the limit on /call is 500 requests a minute
4. Reject webhook and tool requests that don't come from 13.203.39.153, 13.126.9.249 or 13.202.133.53
5. Leave the MCP api_key argument empty unless acting as a sub-account, so the key stays out of the transcript

CueLight

1. Feed 320 float32 samples at 16 kHz mono per push. Resample first, or the call raises an AssertionError
2. Spell the preset as snappy or careful. Any other name silently uses default thresholds
3. Call clear() before a new call or session. Pass reset_on_turn_start=False if your pipeline already resets per turn
4. Pin weights by downloading them to a local folder and passing that path to from_pretrained
5. Don't use it to identify speakers or clone voices. The training data's conditions forbid it