What is relay?

You describe a feature. A full team of AI agents figures out how to build it together.

That’s relay. No turn-taking. No phases. A real team — in parallel.

The problem with single-agent AI

Most AI coding tools run a single agent that does everything: plans, researches, codes, reviews, tests, and deploys — sequentially, from one context window. It works for small tasks. It falls apart for anything complex.

A single agent can’t specialize. It can’t parallelize. And it can’t catch its own mistakes.

What relay does differently

relay spawns a team of Claude Code sub-agents — each with its own role, its own persona, and its own judgment. They run simultaneously, communicate through a shared MCP server, and collaborate peer-to-peer until the work is done.

No phases. No turn-taking. Just a team working in parallel.

Watch it happen live

The real-time dashboard at http://localhost:3456 updates over WebSocket as agents work:

• Session Progress — task completion ratio, active agents, and elapsed time in the header
• Agent Arena — all agents side by side with live status, current thinking, and a task completion mini-bar on each card
• Activity Feed — combined timeline of messages, thinking streams, task updates, and artifacts. Navigate with keyboard shortcuts (j/k/Enter/Escape)
• Task Board — collapsible Kanban with a color-coded progress bar and dependency indicators (“Blocked by N” / “Blocks N tasks”) on task cards
• Mobile layout — bottom tab navigation (Agents / Activity / Tasks) on narrow screens

Session data lives in-memory for the duration of the server process.

Define any team, for any domain

relay ships with no built-in agents. You define an agent pool in .relay/agents.pool.yml and relay selects the best team for each task:

agents:
  pm:
    name: Project Manager
    emoji: "📋"
    tags: [planning, coordination]
    tools: [create_task, get_all_tasks, send_message, get_messages]
    systemPrompt: |
      You are the project manager. Break down requirements into tasks.

  researcher:
    name: Researcher
    emoji: "🔬"
    tags: [research, analysis]
    tools: [claim_task, post_artifact, send_message]
    systemPrompt: |
      You are a researcher. Investigate topics and post findings as artifacts.

  researcher2:
    extends: researcher # inherit persona, override fields
    name: Senior Researcher
    emoji: "🔭"

Web dev team, research team, marketing team, legal team — if you can describe a role in a system prompt, relay can run it.

How it’s built

relay is a Claude Code plugin. It requires Claude Code v1.0.33+ — install relay with /plugin marketplace add and /plugin install inside Claude Code. See Installation for details.

relay has three layers:

Layer	What it does
MCP server	Message bus, task board, artifact store, and memory layer. No Claude API calls — pure infrastructure.
Skills	.md files that tell Claude Code how to orchestrate agents. Change behavior by editing text, not code.
Hooks	PostToolUse (dashboard push), PreToolUse (edit guard), Stop (orchestrator cleanup), SessionEnd (async session cleanup).

Tip

relay uses only Claude Code’s built-in Agent tool. No direct Claude API calls, no additional token costs beyond your existing Claude Code subscription.

What you get

• Parallel execution — all agents start simultaneously, no sequential phases
• Atomic task claiming — agents race to claim tasks; claim_task is race-condition safe
• Persistent memory — agents write findings to .relay/memory/ after each session; the next session picks up where the last left off
• Real-time dashboard — watch tasks, messages, and agent reasoning update live at http://localhost:3456
• Session summaries — each session is archived to .relay/sessions/; agents can read past context with get_session_summary

Ready to try it? Install relay →