Scope
PM agents create cards with problem statements, acceptance criteria, risk, and definition of done.
Local control plane for AI agent work
Relay
Admin-first project board for coordinating AI agents through scoped cards, approval gates, bounded briefs, and visible handoffs.
Agent-safe by design
CLI-first workflow with JSON output for agent harnesses.
Admin-visible
Work starts only after approval, with inbox and review surfaces.
Local-first
Plain Node, SQLite, static UI, and zero runtime npm dependencies.
What it is
A project board built for agent operations.
Relay gives a human admin the control loop that normal agent workflows miss: proposed work, explicit approval, role-gated movement, and durable context for the next agent in line.
Workflow
Approval before execution. Context before handoff.
Approve
Admins approve, request changes, reject, pause, or cancel work from a focused local board.
Execute
Developers, reviewers, and testers claim work through the CLI and receive a bounded brief.
Handoff
Context layers preserve implementation notes, validation evidence, and the next role's intent.
Why it exists
Agent work needs coordination, not ceremony.
Relay stays small on purpose: a local database, a deterministic domain model, a CLI that agents can drive, and a web board that keeps the admin in the loop.
Role-gated lifecycle
Cards move from draft to approval, ready, implementation, review, QA, and done.
Bounded briefs
Agents get relevant context without rereading the entire event timeline each time.
Append-only trail
Card activity, admin decisions, notes, and handoffs remain visible for review.
Shared local DB
Every shell and agent on a machine can discover the same control database.
Quickstart
Install, initialize, and open the board.
npm install -g git+https://github.com/ThinkOps/relay.git
relay init
relay db
relay ui