Overview

LiveGrid is a stack for real-time voice and video calls that software can join, run and transcribe — on infrastructure you rent and control rather than on somebody else's account.

A call is a room on a media server you host. Everything that takes part in one — a browser tab, a phone line, an agent worker, a rendered face — joins that room as a participant holding a signed token, and publishes or subscribes to audio, video and data tracks. That is the whole model. The rest of these pages are the details of putting something into a room and finding out what happened in it.

What you can build with it#

• A voice agent that answers. Audio in, audio out, a transcript afterwards. Call modes
• An agent with a face. The same call with a rendered avatar published as a video track. Avatars
• A console over your own call history. Every turn, every tool the agent fired, every latency. Transcripts
• Calls placed by other software. Seven MCP tools, stateless, over JSON-RPC. MCP tools

Four ways in#

LiveGrid is callable from four surfaces, and they are not alternatives to each other — they sit at different distances from the media.

For coding agents#

Every page here is available as markdown: append .md to any documentation URL. There is a machine-readable index at /docs/llms.txt, and one per section.

LiveGrid is itself an MCP server, so an agent can read these docs and then place a call without an SDK. See MCP tools.

What this is built on#

The media server is LiveKit (opens in a new tab), open source, running on a box you rent. LiveGrid is the layer above it: token policy, region selection, agent dispatch, avatar enrolment and consent, call recording, and the tenancy that keeps one customer's calls away from another's.

If you want LiveKit's own reference for the SFU, the client SDKs or the wire protocol, read their documentation (opens in a new tab) — this site does not restate it.

Where to go next#