Install the runner on any machine, pair it with an 8-character code, and let it dispatch agents from the queue.
Auto-run runs experiments without you at the keyboard. A paired machine polls the brain, launches your coding agents non-interactively, and mirrors a redacted excerpt of each job so the Auto-run page shows what’s happening. It works on a laptop or a remote box, and it is independent of the agent plugin you use interactively.
curl -fsSL https://rapidreview.io/merv/runner/install.sh | sh
The installer needs Python 3.11+ and Git. It installs under ~/.merv, starts the runner, and prints an 8-character pairing code such as 7Q2K-M4B9. Rerun it any time to update.
2
Enter the code in the UI
Go to Auto-run and choose Pair a machine. Approving the code registers a project key labelled auto-run · <hostname>, and the runner starts on its next poll.
The Auto-run page before any machine is paired: enter the code the runner printed.
The machine mints its own key and sends only its digest. No browser reaches the machine, no port is forwarded, and nothing is pasted but the code. If you revoke the key later, the runner stops with a re-pair instruction.
3
Enable agents and a repository
On the machine — executables never leave it:
CLI=$HOME/.merv/bin/merv-client$CLI agent codex --enable --command codex --parallelism 2$CLI agent claude --enable --command claude --model opus --effort high$CLI workspace --repository /path/to/repo --strategy git_worktree$CLI harness # installs Merv skills, checks each agent is ready
The same settings are editable in the UI’s machine drawer: enabled, model, effort, and parallelism per platform, plus the repository, worktree root, and base ref. Changes apply on the next heartbeat and show as Settings pending until then.
4
Check
The machine appears under Machines with its agent inventory. Having no agents enabled yet is a valid state.
The brain keeps only what the UI needs: the non-secret heartbeat inventory, the desired tuning per platform, a settings version, and a bounded, redacted excerpt per job — the last events, the tail of stderr, secrets masked.Everything else stays on the machine under ~/.merv/: client.json (executable argv and custom agents), agent-runner.key and agent-runner.secret, agent-sessions.json, and the full traces in agent-traces/<session-id>/.
Native adapters exist for Codex, Claude Code, Gemini CLI, Cursor Agent, OpenCode, GitHub Copilot CLI, Qwen Code, and Hermes Agent. Anything else runs through the command adapter, which reads the instruction on stdin and emits JSONL on stdout. Aider is unsupported because it produces no complete structured trace.Each child process receives only MERV_AGENT_SESSION_KEY; the runner strips its own credentials first. Codex and Claude Code get an isolated MCP configuration, and the others call tools through merv-client call. Because that hides your own plugin from the child, the runner installs the Merv skills itself and links them into each workspace.
Worktrees are required. The runner keeps a private bare repository with no remote — Merv never pushes to yours — and one persistent branch per experiment under ~/.merv/worktrees, reused across sessions. Worktrees prevent Git collisions between agents; they don’t stop a hostile agent from reading another’s files. Use containers or separate OS users for that.
curl -fsSL https://rapidreview.io/merv/runner/install.sh | sh -s -- --install-only
For CI, create a project key under Settings → MCP keys, export it as MERV_MCP_KEY (not through shell history), and start the runner with $HOME/.merv/bin/merv-agent-runner --project proj_123. On a self-hosted brain, point the client at it first: merv-client configure --control-url https://your-brain.example.com.
Merv
Nisa
