Built a near-0 latency service for phone calling for agents. Would love feedback!
Hello fellow SideProjectors!
A friend (IsN4n) and I have been working with business clients with operational flows that need outbound phone calling. Think:
- initial screening call to qualify a customer based on an intake form,
- calling to confirm they would attend the appointment on the day of appointment,
- navigating an IVR to get to a human, etc.
These calls are too unstructured and too contextual for static or rag based voice agents that platforms like VAPI, Retell, Bland provide, yet repetitive enough to benefit from automation. Additionally, the latency is pretty high due to unnecessary speech to text and text to speech steps in all platforms.
So we built our own full-duplex MCP & extracted it as cocall.ai for all.
The core concept:
- Speech-to-Speech: Uses a native speech-to-speech model for full-duplex conversation, rather than a cascade of STT -> LLM -> TTS which is slower. This makes the latency so low, we needed to add an artificial delay so we don’t interrupt the person on the other side often.
- Contextual Pausing: If the agent hits a question it can't answer mid-call, it pings you for the info instead of guessing while continuing natural conversation with the recipient.
- Live Dashboard: You can view real-time transcripts of multiple calls at once and hit a "takeover" button to jump into any call yourself.
- Verified Caller ID: Requires phone number verification so the recipient sees your actual number as the caller.
- IVR Navigation: The agent can handle the "press 1 for sales" loops and hold-lines, only transferring to you once a human is on the line.
The current version works as a connector for Claude or any other MCP-compatible host. You can listen to calls live, download recordings, and manage the "escalation contract" (how the agent handles failure).
Honest feedback welcome, specifically on the integration with Claude, Codex, Radish etc.