Quick start

Two commands. That’s all it takes to go from zero to a full team of AI agents working through your feature in parallel.

Run a task

Describe what you want built:
```
/relay:relay "Add a shopping cart feature"
```
On first run, relay auto-generates a project-specific agent pool (.relay/agents.pool.yml) by analyzing your codebase — tech stack, frameworks, test runners, and more. You can customize this file anytime.

All agents start simultaneously. Each one claims tasks, communicates through MCP tools, and collaborates peer-to-peer — no orchestrator bottleneck. Watch it happen live at the dashboard URL shown when the session starts.

Commit .relay/memory/ to git after the first session. Your whole team shares the same agent context.
Check the dashboard (optional)

Open the real-time dashboard while agents are running:
```
http://localhost:3456   # default — auto-selects 3457–3465 if already in use
```
Session progress in the header, Agent Arena with task completion bars, Activity Feed with keyboard navigation, and Task Board with dependency indicators — all updating live as work happens.

Invoke a single agent when needed

Need one specific perspective without spinning up the whole team?

/relay:agent fe "Refactor the CartItem component"
/relay:agent be "Security review the user auth API"

What a real session looks like

User: /relay:relay "Implement a login feature"

[PM]       → breaks down login into 5 tasks: UI, API, screen spec, event design, tests
             → assigns to Designer, DA, FE, BE

[Designer] → writes login screen spec → post_artifact("login-screen-spec")
[DA]       → designs login_success / login_fail events → post_artifact("login-events")

[FE]       → reads Designer spec → builds login UI → post_artifact("fe-pr")
[BE]       → designs auth API contract → shares with FE → builds implementation → post_artifact("be-pr")

[FE2]      → reads fe-pr → submit_review (approved with comments)
[BE2]      → reads be-pr → submit_review (approved)

[QA]       → writes test scenarios → creates bug tickets → dev team fixes

[Deployer] → QA signs off → deploys → notifies team

User: Login feature is live.

The whole flow is event-driven. Agents react to each other’s messages and artifacts — no waiting in line.