Callhome is an open-source, self-hosted voice-call stack for AI companions. It supports outbound calls, voicemail, do-not-disturb, and emotion-aware signaling with a browser-based UI and a server-side gateway.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
An open-source voice-call stack for AI companions. It enables a self-hosted setup where a companion can call you, hang up gently, and leave voicemails. It includes features for emotion sensing, tone cues, and configurable behavior in a browser-based (PWA) client and a server-side gateway.
How it works
The architecture includes a Browser (PWA) client and a Server (self-hosted) with components for audio processing and signaling:
- VAD + rec, call UI in the browser
- SenseVoice + librosa for transcription, emotion tagging, and tone cues
- A gateway layer that handles markers and invites and routes to a local LLM and TTS providers
- TTS options (e.g., ElevenLabs with user-provided keys) The project currently documents a workflow where speech-to-text, emotion tagging, and tone cues are produced on the server, and text + emotion + tone are streamed to a gateway which interacts with an LLM and TTS to generate responses.
Getting started
The repository includes a runnable stt-service directory with SenseVoice + librosa in one endpoint for transcription, emotion, and tone cues. It also provides docs/PROTOCOL.md with the full marker and invite protocol (dial, hangup, DND, voicemail, escalation, call records). The gateway-reference directory contains annotated production extracts of the marker layer. Config is intended to reside in a separate config (persona, memory, keys) rather than in code, emphasizing a clone-and-configure approach to set who lives in the house.
Note: The README excerpt provided does not include explicit installation commands or setup steps. See the docs directories in the repository for protocol and runnable services.
Recent releases
RELEASES (latest 0):
- none
Traction
Stars: 18 Forks: 4 Open issues: 0
Behind the repo
Built by Elle & Matt; co-authored by the companion it was built for.
Caveats
License: MIT Created: 2026-07-11 Last push: 2026-07-12 Language: HTML Description: An open-source voice-call stack for AI companions. Your companion can call you, hang up gently, and leave voicemails when they miss you. Self-hosted. Your voice never leaves your server. Notes: The README mentions self-hosted infrastructure, and references to external services (ELEvenLabs) with the caveat to bring your own key. The project relies on components like SenseVoice, librosa, and external TTS services; weights are not distributed here.