rc-agent runs on your own machine: it holds the profiles, starts and supervises the CLI servers, and connects out to the hub. Nothing ever connects to your machine.
rcctl is the optional command-line client for the hub; the agent alone is enough to get started.
Every release ships both binaries for six platforms. They are static, single-file programs: there is no installer. Each release also publishes a SHA256SUMS file to verify them.
From an empty machine to an agent that comes back by itself after a reboot.
The agent starts sessions of an agent CLI; today that CLI is Claude Code, and it has to be installed and on your PATH.
A profile is one CLI identity: its own configuration home, its own login, its own working directories. For each profile, log in once by hand. The agent never sees the credentials.
Pick your platform above, verify the checksum, then make it executable and put it somewhere stable on your PATH.
It prints the address of its local panel, which binds to the loopback interface only and never renders a secret value.
Open the Pairing tab, enter the service URL and approve the code from a signed-in device: the mobile app, the web panel, or rcctl pair approve.
Add one profile per CLI identity, with its home, its base directory for new sessions and an allowlist of extra working directories.
So the agent starts at login and comes back after a crash: a systemd user unit on Linux, a launchd agent on macOS, a service on Windows.
A machine can also run OpenAI's Codex CLI, and its sessions are picked up in the ChatGPT app instead of the Claude app. rc-agent runs one Codex daemon per profile.