Quickstart · v0.6.70
Two lines of shell. Everything else is the session.
Install it, start it, answer five prompts once. From then on you type tasks into Relay’s own interface — there are no commands to learn.
Step 1
Install and start Relay
macOS and Linux, or WSL2 on Windows. No sudo, nothing outside your home directory.
Then, from inside a git repository:
That is the last shell command on this page. The first run walks its five steps itself — there is no separate login command and no separate connect command:
That is a real capture, and it stops where the capture stops: the next thing is your browser. Signing in binds this machine’s signing key — what every later verdict is signed against. The remaining steps need no typing: detect probes the agent CLIs on your PATH, agents asks which to use and in what failover order, models picks the model for each, and ready sets up the repository you are standing in.
Run it in a terminal, not through your agent. The last step asks you to vouch for this machine’s key, and that is the one decision an agent must never make for itself — a machine that approves its own signing key makes every verdict after it worthless.
Step 2
Type your first task
The composer says Type a request or / for commands. Anything you type that is not a / is the task:
Relay maps your words to real files, decides what done means, and hands the work to your agents — each turn in its own isolated worktree, with failover to the next agent if one stalls. You watch it happen in the session: the files it found, the plan, the agents working, then the checks.
You never run the tests yourself. Relay re-runs them, measures which changed lines actually executed, and signs what it measured — nothing an agent says is ever an input to the verdict.
Step 3
Read the verdict
Every run ends the same way: what it understood, whose plan it followed, who built it, who measured it, and what it did and did not establish — then the verdict.
PASS means signed evidence covers every changed line. BLOCK is a real finding, named. NEEDS CLARIFICATION means it could not find what you asked about. NEEDS_HUMAN is the rare thing only you can answer — and the loop stops and waits rather than guessing.
Step 4
Go again — or follow its recommendation
A BLOCK is not a dead end: the run ends with a recommendation, and you can queue it as the next task without retyping anything:
Or just type the next task. The session keeps your agents, your models and what Relay learned about the repository — lessons survive between runs, so it does not learn the same thing twice.
Step 5
Inside the session
The whole surface, from the session’s own dispatch. Typing / alone lists these:
| You type | What happens |
|---|---|
anything else | That is the task — Relay drives it |
/agents | The board: your chain, models, repo, last verdict |
/model | Change an agent’s model (tab also switches) |
/follow | Queue the last run’s recommendation as the next task |
/connect | Re-run agent detection and connection |
exit | Leave the session |
Then
Where to go from here
- The loop — every state a drive can end in, and the exit code for each.
- Measured coverage — what “covered” means here, and the one policy field worth knowing about.
- CI — the merge gate, with or without a stored key.
- The full reference — everything above and the rest of it.
A PASS means the evidence exists and covers the change. It is not a claim that your code is correct, and § 19 is the list of things it deliberately does not prove.