vibelive

Multiplayer for agentic coding terminals — shared Claude Code/Codex/Gemini sessions with live Figma-style cursors, presence, and in-terminal chat.

Part of the Vibe Suite — companion tools for agentic coding CLIs (Claude Code, Codex, Gemini, Grok/pi, Kimi). Ships as CLI + npm package + MCP server.

Local-first: runs on your own machine. v0 is a host-authoritative local/LAN session: one machine hosts a wrapped agent and owns the truth; peers join over a local WebSocket. Full end-to-end-encrypted relay fan-out to ~1,000 participants is on the roadmap (see docs/tech-spec.md). Session sharing is gated by the consent ledger in @pooriaarab/vibe-core (scope share:session), and is revocable.

Install

npm install -g vibelive
# or use it as a library / MCP server without the global install:
npm install vibelive

Requires Node ≥ 18.

Quick start

Share a wrapped agent (anything you'd run in a terminal) and print a join URL:

vibelive host -- claude          # share a Claude Code session
vibelive host -- python -i       # share a REPL
vibelive host --port 4474 --name ada -- claude

The host prints something like:

vibelive host ready — sharing claude
  join: ws://localhost:54157
  lan:  ws://10.0.0.179:54157
  you are the driver. /release to hand off, /drive to take back, /quit to end.

From another terminal (or another machine on the LAN):

vibelive join ws://localhost:54157 --name ada

The joiner sees the live agent output plus a presence/chat line. Slash commands while joined:

• /drive — request the write token (queued FIFO behind other drivers)
• /release — relinquish the token
• /type <text> — send input to the wrapped agent (driver only)
• /quit — leave the session

Everything else you type is sent as chat.

Who can write to the agent?

Exactly one participant — the driver — holds the write token at any time (see src/arbitration.ts). Everyone else always has read + chat + cursor; only agent-write is arbitrated, so two people never interleave garbage into one stdin. The host user starts as the driver.

MCP server

Run vibelive as a Model Context Protocol server over stdio:

vibelive mcp

It exposes two tools an agent can call:

Example (Claude Code / any MCP client config):

{
  "mcpServers": {
    "vibelive": { "command": "vibelive", "args": ["mcp"] }
  }
}

Library

npm install vibelive

import { createHost, createRelay, joinSession, WriteArbiter } from 'vibelive';

// Host-authoritative session on an ephemeral port.
const host = createHost({ command: ['claude'] });
const relay = await createRelay({ port: 0, hostHandle: host, initialDriver: 'host' });
console.log(relay.url); // ws://localhost:<port>

const client = joinSession({ url: relay.url, name: 'ada' });
client.onOutput((text) => process.stdout.write(text));
client.requestControl(); // ask to drive

How it works (v0)

Three channels over one WebSocket, each with the guarantees its data needs (details in docs/tech-spec.md):

1. Agent output — ordered, sequence-numbered append-log; the host is the sole author; late joiners get snapshot + tail.
2. Presence / cursors — ephemeral, high-frequency, lossy is fine (coalesced/interpolated client-side; the relay forwards for v0).
3. Chat + control — reliable, ordered; agent-write is mediated by the WriteArbiter so there is never more than one driver.

The correctness-critical piece is src/arbitration.ts — a pure, fully unit-tested state machine enforcing "never two concurrent agent-writers," FIFO turn-taking, and *"release when not driver is a no-o

... View full README on GitHub