voicegateway 0.15.0

Cost tracking and reconciliation for LiveKit voice agents: modality-aware unit accounting (audio-minutes, tokens, characters) backed by voice-prices.

Docs · Quick start · Dashboard · Fleet collector · MCP

from livekit.agents import AgentSession
from voicegateway import inference          # the only line that changed

session = AgentSession(
    stt=inference.STT("deepgram/nova-3"),
    llm=inference.LLM("openai/gpt-4o-mini"),
    tts=inference.TTS("cartesia/sonic-3"),
)
# every call logged: provider, model, tokens, $cost, latency, session_id

A drop-in cost and quality observability layer for LiveKit Agents. Use voicegateway.inference for VoiceGateway-managed plugins, or voicegateway.attach(session) to observe any existing LiveKit STT/LLM/TTS plugin in place. Modality-aware unit accounting (audio-minutes, tokens, characters) with LLM, STT, and TTS prices from voice-prices, reconciled against your real provider invoices with one command. Self-hosted. Your keys. No data leaves your infra.

Why VoiceGateway

Voice AI vendors hide three numbers. VoiceGateway exposes them, per call.

• Is it working? Latency p50/p95 across the STT → LLM → TTS loop, interruption rate, dead air, talk-over: the metrics text stacks never have to think about.
• What does it cost? STT bills by audio seconds, LLM by tokens, TTS by characters. Every call is broken down by modality and totaled to the cent. voicegw reconcile checks recorded numbers against your actual provider invoices.
• How do I make it cheaper? Route by combined STT + LLM + TTS latency budget across providers, switch models per call type, attribute cost per tenant so agency clients see only their own usage.

Building a text-only LLM app with no voice component? LiteLLM is the better fit. See the decision tree.

Features

Capability	What it gives you
LiveKit Cloud parity	Drop-in for livekit.agents.inference. Your keys, your config
Voice-conversation metrics	Per-minute cost, latency p50/p95, interruptions, dead air, talk-over
Conversation replay	Scrub any past call: STT chunks, LLM tokens, TTS frames with timing and cost
Terminal UI	voicegw tui opens a vim-key Textual UI for SSH-in inspection
Multi-tenant attribution	Per-tenant cost, scoped API keys per team, agency-ready
Cross-modality routing	Route by combined latency budget, per-project rosters, white-label branding
Voice-specific guardrails	Real-time PII detection in STT, prompt-injection detection, compliance hooks
Daemon-first onboarding	Curl-bash install, OS daemon, five-question wizard, voicegw doctor
Fleet collector	One-line installer. N agents push to one collector. Slice costs by agent, project, tenant

Full release history: CHANGELOG.md.

Quick start

# Single node: local SQLite + the dashboard at http://localhost:8080
pip install "voicegateway[cloud,dashboard]"
voicegw init && voicegw serve

Add the three inference lines from the snippet above to your agent and every call is tracked. Provider plugins install modularly (pip install "voicegateway[deepgram,openai,cartesia]"). The full extras matrix, the zero-install uvx path, and the OS daemon installer (LaunchAgent / systemd / Scheduled Task) are in the get-started docs. Python 3.11+.

The dashboard

A self-hosted web UI at http://localhost:8080. Bundled. No SaaS account. No data leaves your stack.

_{A 7-day spend / requests trend, per-provider cost breakdowns, a live request log, and one-key light/dark.}

Overview with a 7-day spend and requests trend, Costs (per provider / model / project / tenant, plus latency p50/p95), Sessions (replay, routing decisions, budget overruns), Logs, Agents, and Settings. One-key light/dark theme. White-label it per project: upload a logo, set an accent color and product name, and the whole UI re-skins.

Fleet collector

Run one shared collector on your VPS. Every agent on your fleet pushes telemetry to it: one dashboard, one cost view, across all of them.

curl -fsSL https://voicegateway.dev/collector.sh | bash

The script installs Docker if needed, generates and persists secrets, pins the image version, and health-checks the container before returning. Point your agents at it:

from voicegateway.services.sinks import RemoteCollectorSink

sink = RemoteCollectorSink(
    collector_url="https://collector.example.com",
    api_key="<your-ingest-key>",
)

SQLite and Postgres backends, plus HTTPS via Caddy: fleet collector docs →

Manage from your coding agent (MCP)

VoiceGateway ships a first-class Model Context Protocol server. Claude Code, Cursor, Codex, and Cline configure providers, create projects, check costs, and tail logs through natural language.

pipx inject voicegateway "voicegateway[mcp]"
claude mcp add voicegateway --command "voicegw mcp --transport stdio"

22 tools across observability, providers, models, and projects. Destructive ops (delete_*) require explicit confirm=True after a preview. Remote HTTP/SSE transport with bearer auth and the full tool list: MCP reference →

Supported providers

11 providers across cloud and local. Mix and match per call.

Modality	Cloud	Local
STT	Deepgram, OpenAI Whisper, AssemblyAI, Groq	faster-whisper
LLM	OpenAI, Anthropic, Groq	Ollama (any compatible)
TTS	Cartesia, ElevenLabs, Deepgram Aura-2, OpenAI	Kokoro, Piper
VAD *	Silero	Silero
Turn detector *	LiveKit MultilingualModel	None

* Configured directly on the LiveKit AgentSession, not wrapped by VoiceGateway. The 11-provider count covers STT, LLM, and TTS.

Per-model IDs: configuration/providers. Adding a provider takes about 10 steps: contributing/adding-a-provider.

Architecture

flowchart TB
    A[LiveKit Agent] --> B[voicegateway.inference]
    B --> C[Router]
    C --> D[Cloud Providers]
    C --> E[Local Providers]
    B --> F[Middleware Pipeline]
    F --> F1[Cost Tracker]
    F --> F2[Latency Monitor]
    F --> F3[Guardrails]
    F --> F4[Multi-tenant Attribution]
    F --> G[(SQLite · encrypted)]
    G --> H[Dashboard UI]
    G --> I[MCP Server]
    I --> J[Claude Code · Cursor · Codex]

Async throughout. Modular provider installs pull only what you use. YAML config with ${ENV_VAR} substitution. SQLite at the bottom for portability, encrypted with Fernet at rest. Architecture deep dive →

Deploy

Docker Compose (Postgres + collector)

services:
  postgres:
    image: postgres:16-alpine
    environment:
      POSTGRES_USER: voicegw
      POSTGRES_PASSWORD: ${VOICEGW_PG_PASSWORD}
      POSTGRES_DB: voicegw
    volumes:
      - voicegw-pgdata:/var/lib/postgresql/data
    restart: unless-stopped

  collector:
    image: mahimairaja/voicegateway:0.9.2
    ports:
      - "8080:8080"
    environment:
      VOICEGW_DB_URL: postgresql+asyncpg://voicegw:${VOICEGW_PG_PASSWORD}@postgres/voicegw
    volumes:
      - ./voicegw.yaml:/app/voicegw.yaml:ro
    depends_on: [postgres]
    restart: unless-stopped

volumes:
  voicegw-pgdata:

docker compose up -d

For production, the fleet collector installer handles secrets, image pinning, and health checks for you.

Contributing

git clone https://github.com/mahimailabs/voicegateway
cd voicegateway
pip install -e ".[all,dashboard,mcp,dev]"
pytest

Read CONTRIBUTING.md and CODE_OF_CONDUCT.md before opening a PR. Security issues go through the disclosure flow in SECURITY.md, not a public issue.

Community

License

MIT. Fork it, ship it.

Built by Mahimai Raja, founder of Mahimai AI, a voice AI company, in public. Standing on LiveKit Agents, FastAPI, Pydantic, and voice-prices.