callsmith

A recipe compiler for voice AI agents. It takes a short MCQ intake (surface, telephony, orchestration, hosting ownership, architecture, realtime model / STT+LLM+TTS, VAD, language, barge-in, audio cleanup, debug depth, voice UX, latency priority, business logic, tools, handoff, safety, deployment), resolves audio-format and operational compatibility, computes the interruption/turn-taking flow, models the latency budget, and writes an agent handoff packet: callsmith.recipe.md + callsmith.lock.json + .callsmith/context/*. A coding agent then builds the full system from that packet — using framework-native code (LiveKit Agents, Pipecat PipelineWorker, or custom FastAPI) — without rediscovering telephony potholes.

The hard part of voice agents is not the model. It is the audio bridge (μ-law 8 kHz ↔ PCM 16/24 kHz), turn/interruption lifecycle, VAD configuration, echo/noise cleanup ownership, and per-provider streaming quirks. callsmith carries that knowledge so the implementer does not have to.

Setup check

Run this before anything else. If the CLI is missing, install it.

command -v callsmith >/dev/null 2>&1 && echo "Installed" || npm install -g github:Nachi-Kulkarni/Callsmith_skills

In a checkout of the callsmith repo itself, the CLI runs without a global install:

node bin/callsmith.mjs --help