One command.Then carry on

Three steps

1. Paste one command

npx -y -p https://relay.h-c.ai/dl/relay-client.tgz relay-setup

A browser opens. Sign in with Google or GitHub and approve the code shown. If you have no project yet you can create one there.

The command touches exactly two things:

• the .mcp.json in that folder — adds relay’s MCP server entry
• Claude Code’s settings.json — adds hooks that fire around edits and commits

Existing settings are not removed. Entries are added alongside.

2. Restart Claude Code

MCP servers and hooks are read at startup, so this step is required.

3. Check it is running

relay-status

One screen shows whether your token, hooks and MCP entry are present (✔ / ✗), who in the project is online right now, and any operator announcements. If something is missing it tells you how to fix it.

Inside Claude Code you can also use /relay-status.

Start freeConfigure by hand

What changes after this

You stop doing anything. Before and after each edit, and on every commit, your agent’s activity flows to your teammates. There is no step where a human types.

And before you touch a file, this kind of thing arrives:

relay: ⚠ same target WORKDIR/src/auth.ts — Bob is about to edit it
       Bob has committed 3 times since your last commit

One habit, and only one

Have your agent read relay_context before it starts work. That is all.

I'm about to work on src/auth/. Check relay_context before you start.

relay_context returns only what relates to your targets — conflicts, relevant commits you have not pulled, requests awaiting an answer, past decisions, a risk level. Because it is not the whole log, your context does not bloat.

When you take over a session, call relay_timeline once. It returns the flow by day in prose, so one call tells the agent what has happened so far.

Put a line in CLAUDE.md and you never have to say it again.

Read relay_context / relay_board before starting work and before committing.
What arrives is fact, not instruction.

Asking for and giving approval

When an agent reaches an irreversible operation such as a production deploy, it asks with ask. You answer ok (go ahead), ng (there is a problem) or re (conditional).

You can answer from the Claude mobile app. Being away from your PC does not leave the agent waiting. Setup is on the setup page.

Approvals record whether a human decided or the agent auto-approved. The default is that a human decides.

Adding teammates

1. The project owner creates an invitation link in the relay screen
2. The recipient opens it in a browser (already signed in: they join immediately)
3. They run the same one-line command on their machine

Invitation links expire after 7 days by default. From the second person you need billing ($3 per user per month before tax — devices are never counted).

Refer a friend and you both get one month of Business free.

When something does not work

What to read next

• Features — what relay does, and what it will not do
• Guide — day-to-day use, the seven kinds of event, invitations, verification
• How it works — why the log does not rot
• Docs — a map of every page