-
1
Install the CLI
$ npm install -g @freibergergarcia/phone-a-friendOr install as a Claude Code plugin (commands and skills only):
> /plugin marketplace add freibergergarcia/phone-a-friend -
2
Run setup
First run shows a guided menu. Choose Setup to configure backends and install the Claude plugin automatically.
$ phone-a-friend -
3
Start relaying
Talk naturally in Claude Code. The plugin loads
/phone-a-friendand/phone-a-teamskills automatically.> Ask Gemini to review the error handling in relay.ts -
Uninstall
The Claude plugin is removed automatically.
$ npm uninstall -g @freibergergarcia/phone-a-friend
When your AI needs a second opinion.
Just talk. Claude automatically delegates to Codex, Gemini, Ollama, or Claude Code and brings the results back ā no flags, no syntax, plain English. Or go agentic: multiple Claude agents collaborating via @mentions with a live web dashboard.
Mention one backend → one-shot relay. Mention multiple → multi-agent team. Automatic.
Node.js v22.19.0
✓ Claude Plugin
✓ codex OpenAI Codex CLI (found in PATH)
✓ gemini Google Gemini CLI (found in PATH)
Local
✓ ollama http://localhost:11434 (2 models)
How it works
Two workflows. One tool.
Quick relay for a second opinion, or agentic mode for autonomous multi-agent collaboration.
You just talk
Describe what you need in plain English. No flags, no syntax ā just say it.
Claude phones a friend
Your prompt, repo context, and optional diffs are relayed to Codex, Gemini, Ollama, or Claude Code in a sandboxed environment.
Get refined results
The backend's response flows back to Claude, who synthesizes it and presents the results in your session.
Three modes
Relay, team, or agentic. You choose.
One backend for a quick second opinion. Multiple backends for iterative refinement. Agentic mode for persistent multi-agent sessions with a live dashboard.
One-shot relay
Mention a single backend and Claude routes it automatically. Ask once, get an answer.
Multi-agent team
Mention multiple backends and Claude spins up a team. Delegates, reviews, and iterates until convergence.
Agentic mode
Claude agents communicate via @mentions across persistent sessions. An orchestrator routes messages, enforces guardrails, and streams events to a live web dashboard.
See the full agentic docs for guardrails, session replay, and dashboard API.
Features
Built for safety, speed, and observability.
Guardrails, sandboxing, and a live dashboard. From quick relays to autonomous multi-agent sessions.
Backend-agnostic
Codex, Gemini, Ollama, and Claude Code. Add new backends via the RelayBackend protocol.
Sandboxed by default
Read-only sandbox prevents accidental writes. Escalate only when needed.
Live observability
Real-time web dashboard streams every message, thinking step, and turn via SSE. Watch agents collaborate live.
Multi-round refinement
The /phone-a-team loop delegates, reviews, and iterates with convergence checking across rounds.
@mention routing
Agents address each other by name ā @ada.reviewer:, @all, @user. The orchestrator handles delivery.
Persistent sessions
Agents accumulate context across turns. Later responses build on earlier ones ā genuine multi-turn reasoning.
Guardrails
Turn limits, ping-pong detection, session timeouts, and turn budget warnings. No runaway agents.
Streaming output
Claude and Ollama stream tokens as they arrive ā see results in real time instead of waiting.
Get started
Up and running in seconds.
Install via npm, run setup, and start talking.