FAQ

How is this different from OrbStack or Docker Desktop?

OrbStack and Docker Desktop manage the Docker / Linux-VM world. Runbay is a GUI for Apple's own container runtime — the one announced at WWDC 2025 that reached CLI 1.0.0 in June 2026. Each container runs in its own lightweight VM with a dedicated IP and sub-second startup, with no shared kernel. Apple ships that runtime CLI-only; Runbay adds the GUI, plus three things none of the alternatives have for this runtime: compose-style Stacks, a restart Supervisor, and a two-way MCP bridge for coding agents. It is not a Docker wrapper, and it does not expose a Docker socket.

Does it work on macOS 15, or do I need macOS 26?

It runs on both, but macOS 26 Tahoe is the full experience. Apple officially supports the container CLI only on macOS 26, where you get custom networks and container-to-container networking. On macOS 15 you lose container-to-container networking and container network create — that's an Apple constraint, not ours, and the app tells you so honestly in-app. Apple Silicon is required either way; the framework does not support Intel.

Is my data local? What leaves my machine?

Everything runs on your Mac. Containers, logs, stats, and the SQLite cache are all local. The MCP server listens on a local Unix socket or localhost TCP only, rate-limited. There is no telemetry backend, no hosted agent compute, and no config, license, or sync server on our side. The only outbound network call in normal use is license validation to Polar (using a public organization ID) — see licensing.

What is MCP, and why does it matter here?

MCP (the Model Context Protocol) is the open standard coding agents use to discover and call tools. Runbay is an MCP server: every running container automatically becomes a set of tools (logs__, stats__, exec__, and db_query__ for databases), plus live resources and lifecycle events. Your agent can read logs, check stats, run a read-only query, or spin up a disposable sandbox without you writing any glue. Full detail in the agent setup guide.

Can my coding agent touch my host filesystem or credentials?

Not by default. Agent Sandboxes are disposable container machine VMs, and the only host-filesystem lever the CLI offers is --home-mount none|ro|rw. The default is Isolated (none) — your $HOME, SSH keys, and cloud tokens stay invisible. Choosing ro or rw exposes your home directory, and the app warns you in red before you do.

What are the license terms?

Free for personal and non-commercial use, with every feature included. Commercial use at a company requires an ~$8/user/month subscription (honor-system license key, like OrbStack). See licensing for how activation and offline grace work.

What doesn't work yet (honest limitations)?

Honesty is the point — here's the real list.

  • macOS 15 has no container-to-container networking and no container network create. That's an Apple constraint; the full experience needs macOS 26.
  • The app can't restart a fully-dead engine by itself. Under the App Sandbox, launchd refuses job submission from a sandboxed process, so the watchdog degrades to a guided "Restart in Terminal" button and self-heals on the next poll. The fully-automatic fix needs a non-sandboxed helper — the detail is in the supervisor doc (decision D017).
  • Image builds are slow upstream. Apple's container build is currently ~4–5× slower than OrbStack / Docker Desktop, and Dockerfiles at or above ~16 KB can fail. Our Dockerfile build UI is deliberately unshipped until we can preflight those traps and frame the performance honestly.
  • linux/amd64 emulation can segfault JIT-heavy workloads (Node, .NET, MSSQL). Prefer arm64 images.
  • No Docker socket. Apple's runtime exposes none, so Docker-socket tools — including Testcontainers — don't work today. A bridge is under evaluation.
  • No pause / unpause / attach. These aren't in CLI 1.0.0; attach flows use exec -it or start -a -i.
  • Volume data doesn't migrate automatically from Docker/OrbStack. Images re-pull; volume data you move yourself. See migration.
  • MCP has no TLS yet. It's localhost-only and rate-limited (100/min, 1000/hr per client) — keep it to your machine.

Why should I trust an honest limitations list over a polished one?

Because it's the whole pitch. Every claim on this site traces to the project's own tracking docs, and the capability matrix in the README leads with what doesn't work. If something here is wrong or out of date, that's a bug we want to fix — not a line we're trying to hold.