Security
This release fixes several security problems, most of them reported to us privately. Update now. Details are in the advisories on the repository's Security tab.
- An image can no longer reach files on your Mac that you didn't share. The Linux VM is now given
/Users,/Volumesand/private(as Docker Desktop shares by default), not the whole disk, and mount settings that come from an image are checked more strictly. - A container's working directory always stays inside the container, and so does the folder Kiri makes for it when it is missing.
- A new volume is filled without following links inside it, so a container sharing the volume can't send the copy somewhere else.
- A malformed DNS query from a container can no longer stop the engine.
- Mac: containers can't reach other containers through the domain router (
*.kiri.localhost), and its "not found" page lists your running containers only when asked by one of your own machine's names. - Seccomp profiles are applied, and the default filter matches Docker's.
- Linux and WSL: containers are firewalled on kernels before 6.3, and reach the internet there.
- An
--initcontainer's first process keeps none of the engine's open files.
New
- Compose has its own section in the window.
- Install with Homebrew:
brew install --cask libracoder/kiri/kiri. - Every menu in the window can be searched: type to filter fonts and other long lists, with the arrow keys, Enter and Escape.
- Windows: Settings → Import brings images, volumes and containers over from Docker Desktop.
- Loop devices work in containers (
losetup, disk images). - Container files in Finder show volumes and bind mounts.
Fixed and improved
- Compose 5.7 and later attach again:
docker compose upanddocker compose logs -fshow output. - Testcontainers for .NET, MongoDB and LocalStack start: Kiri closes a connection when the client asks it to, answers
execandattachas Docker does, joins a container only to its own network, and a file copied in straight after a start is there. docker execstraight after a start finds the image's users, and a start that fills a new volume no longer holds up every other container.- Saves made on the Mac reach file watchers in containers, so hot reload works in bind-mounted folders.
alwaysandunless-stoppedcontainers come back after the VM crashes, and Kiri restarts on the new version after every update.- A published TCP connection stays open however long it is quiet.
- Pulling several images at once can no longer damage the image list, and a damaged one is repaired. A damaged container list no longer costs containers their logs, files or volumes.
- A named pipe, or a program that never stops writing, no longer holds up the engine.
- Images with hard links to absolute symlinks unpack (Supabase's PostgreSQL image).
- k3d finds its nodes: the name filter is a regular expression, as in Docker.
- Images list newest first, and a missing working directory is made for non-root and read-only containers.
- A tmpfs volume keeps its owner and mode.
- Image update checkers (Watchtower, Arcane) work: Kiri answers
GET /distribution/{name}/json. docker inspectreports a container's port bindings and exposed ports.- Image names with encoded characters resolve (
mongodb%2Fmongodb-atlas-local). docker cpnever mounts a container's root twice, and containers start and commit correctly on kernels with overlayfs index on (Arch, CachyOS).- Signals only reach running processes, and a failed start leaves no DNS name behind.
- Import brings volumes, containers that share a network and digest-pinned images across, leaves a volume in use alone, and says plainly when OrbStack or Docker Desktop isn't running.
- Windows: Docker works again after WSL restarts, launching Kiri at login no longer undoes a Stop, and the window opens whatever the page's size.
- Mac: ⌘C, ⌘V, ⌘X, ⌘A and ⌘Z work in the window's text fields, the traffic lights stay in the strip after switching themes, and opening Kiri from Finder no longer quits silently when a shell profile is a link to nothing.
- Mac: a first-launch step that fails no longer closes Kiri. It waits in the setup checklist with its error and Try again, and the window works meanwhile.
- Linux: an engine restart cleans up containers whose cgroup systemd already removed.
Downloads for macOS, Windows and Linux: https://runkiri.com/releases/