Skip to content
Projects

mahimailabs/voicegateway has 59 GitHub stars and 14 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-18 13:00 UTC. Last push 2026-11-16.

How these counts are measured

What you can do

  • Call inspection: transcript, turns, LiveKit events, the agent's spans and reported usage, in one drawer. Opt-in recordings play beside the transcript.
  • Turn diagnosis: speech to text, turn detection, model first token, tools and voice first audio on the turn's own clock. Overlapping stages are shown as overlapping, never added up.
  • Issues: problems found on many calls become one issue, with the calls as evidence. Resolve, ignore, tag or archive them; a resolved issue that comes back shows as regressed.
  • Every LiveKit layer: media server rooms and participants, SIP trunks and call legs, egress and ingress jobs, and agent workers with jobs, load, CPU and memory.

Measured star growth

in 24h
0
in 7 days
Collecting history
in 15 days
Collecting history
in 1 month
Collecting history

Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.

About

The One Dashboard every Voice Agent needs

AI

README.md

Original

VoiceGateway

One call. The whole story.

The open-source profiler for LiveKit voice agents, from the phone line to the model.

Opening a captured LiveKit call in the VoiceGateway console: the call list, the call's transcript and recording, then the diagnosis of each turn with speech-to-text, turn detection, model and voice timings.

VoiceGateway profiles LiveKit voice calls from the phone line to the model. Every call gets one view: the transcript and recording, each turn's timings, the agent's spans, and the media server, SIP, egress and agent workers around it. You see where a call slowed down or went wrong, and which layer did it.

[!NOTE] V1 preview. Self-host the Go backend and console against your own LiveKit project today. Alert delivery and capacity tests are still being built (roadmap). VoiceGateway Cloud, the same product hosted, opens at voicegateway.dev.

Features

The console Overview: median answer time on a gauge, every turn against the response budget, and the Live wire from callers through the phone bridge and media server to the agent and its models.
  • Call inspection: transcript, turns, LiveKit events, the agent's spans and reported usage, in one drawer. Opt-in recordings play beside the transcript.
  • Turn diagnosis: speech to text, turn detection, model first token, tools and voice first audio on the turn's own clock. Overlapping stages are shown as overlapping, never added up.
  • Issues: problems found on many calls become one issue, with the calls as evidence. Resolve, ignore, tag or archive them; a resolved issue that comes back shows as regressed.
  • Every LiveKit layer: media server rooms and participants, SIP trunks and call legs, egress and ingress jobs, and agent workers with jobs, load, CPU and memory.
  • Live: a project-scoped feed pushed from the worker, with stale and disconnected sources marked as such.
  • For your team and tools: organisations, roles and invites, Google, GitHub or Auth0 sign-in, project API keys, an MCP server for coding agents, signed embeds for your own pages, and a Python SDK for custom call fields.
  • Alerts: seven rules preview on your real calls today; delivery by email and webhook is next.

Unreported measurements stay unknown, never zero. Traces, usage and recordings appear only when the agent sends them.

How it fits

V1 observation boundary: phone/SIP connects through the LiveKit media server to a voice agent and its models. Webhooks, metrics, spans and session uploads reach VoiceGateway outside the audio path.

VoiceGateway stays outside the audio path. The V1 platform joins LiveKit webhooks, server metrics, agent spans and session uploads into one call view. If the profiler is unavailable, the agent's call must carry on. See the architecture and capture design.

Components

VoiceGateway components. LiveKit sends webhooks and metrics, and voice agents send spans and session uploads, to web. The dashboard and MCP read through web's API. web appends to a Redis stream and reads Postgres; worker consumes the stream and writes batches to ClickHouse. An S3 bucket holds opt-in recordings.

One Go binary runs as web (checks keys, takes data in, serves the API and dashboard) and worker (joins events into calls, writes to ClickHouse). A Redis stream is their only link. Postgres holds projects and keys, ClickHouse everything measured, and your own S3 bucket any recordings. make up starts all six. See ARCHITECTURE.md.

Quickstart

Self-host with Docker, git and make. The installer clones the release, starts the six services and runs the preflight checks:

curl -fsSL https://voicegateway.dev/install.sh | bash
cd voicegateway
make user-add [email protected]
make project-add NAME="Production"   # with LIVEKIT_URL, LIVEKIT_API_KEY and LIVEKIT_API_SECRET set

Sign in at localhost:8080. The Connect your agent panel gives the one variable your agent needs, LIVEKIT_OBSERVABILITY_URL, and the webhook URL for LiveKit. Start the agent session with record=True and place a call. Read the self-hosting guide or try the example agent.

Look around without LiveKit

Run the console on captured LiveKit sessions with Node.js 24 and npm:

git clone https://github.com/mahimailabs/voicegateway.git
cd voicegateway
npm --prefix dashboard ci
make dev

Open localhost:5180, choose Calls and open a call.

Inspect a real turn

Diagnosis of a captured call: five turns in 4 minutes 57 seconds, every measured answer over the 1.5 second budget. Turn 2 answers in 3.04 seconds, with a 668 millisecond end-of-turn delay, 685 milliseconds to the model's first token and 1.38 seconds to the voice's first audio.

In this captured call, the caller asks for the status of order 1042. Every measured answer is over the 1.5 s budget, and turn 2 is the slowest at 3.04 s. Its breakdown shows where the time went: 1.38 s to the voice's first audio, against 685 ms to the model's first token and a 668 ms end-of-turn delay. That makes the next step specific: the text-to-speech provider before the prompt.

Read the console docs

Upgrading from the API token

VOICEGATEWAY_API_TOKEN is gone. Create a key per project with voicegateway key create --org <org> --project <project> --name <name> (or in Settings, Access, API keys), send it as Authorization: Bearer <key>, then delete the variable.

The road to V1

V1 is 32 of 39 tasks done; the roadmap issue lists each one and what is left.

  • Self-hosted: runs on your infrastructure today as a preview. Follow Self-hosting.
  • VoiceGateway Cloud: the same product hosted, with self-serve sign-up and checkout. Follow voicegateway.dev.
  • Next: alert evaluation and delivery, more alert channels, the capacity test runner and a 1,000-call load test.

Current progress · Product scope · Decisions

Contribute

Pick an open issue, agree on scope and open a focused PR against main. Start with CONTRIBUTING.md.

From the repository root:

make test
make lint

These run the Go tests (container tests need Docker) and lint, plus the dashboard tests, lint and TypeScript checks. make up starts the six-service stack.

Community and security

Discord · Report a bug · Contributor code of conduct

Please report vulnerabilities privately using SECURITY.md.

License

Free to self-host. Code outside ee/ uses the MIT license; code inside it uses the Enterprise license. VoiceGateway Cloud hosts the same product. Read the boundary.

Sponsors

Sponsored by GitAds