Skip to content

Platform support

Clauster runs on Linux, macOS, and Windows, and the same test suite exercises all three in CI. Most capabilities behave identically everywhere; the matrix below is the single source of truth for what works where and which platform-bound mechanism backs each cell. When a row changes, update it here rather than restating the gap in another doc — the scattered notes elsewhere point back to this matrix.

Legend: ✓ works · ✗ not available — an honest platform gap, not a defect. Bracketed numbers point at the notes below.

Capability matrix

Capability Linux macOS Windows
Standard (Server Mode) bridge ✓ ✓ ✓ [1]
Interactive Session (PTY) bridge ✓ ✓ ✓ [2] [3]
Hosted channel (claustrum) ✓ ✓ ✓ [4]
Dashboard claude login ✓ ✓ ✓ [5]
Config-write CLI (claude mcp / claude plugin) ✓ ✓ ✓ [6]
Per-bridge CPU % / RSS memory [7] ✓ ✓ ✓
Per-bridge disk I/O rate ✓ ✗ [8] ✓
Advisory file locking (config writes + bridge lifecycle) ✓ ✓ ✓ [9]
Owner-only file modes (0o600 / 0o700) ✓ ✓ ✓ [10]
Directory fsync (crash durability) ✓ ✓ ✓ [11]
Service-unit install (install-service) [12] ✓ ✓ ✓

Notes

  1. Windows spawns the bridge with CREATE_NEW_PROCESS_GROUP and stops it via CTRL_BREAK_EVENT; POSIX uses start_new_session + SIGINT.
  2. POSIX uses pty.openpty + termios; Windows drives a ConPTY keeper via pywinpty (#903), which needs the pty extra — without it launch_mode: pty falls back to Server Mode. The extra pulls pywinpty on Windows and pyte everywhere (the live pty-screen terminal view needs it on every OS). Install it with pip install 'clauster[pty]' on a package install, or clauster deps install pty on the standalone binary — the extra is intentionally not bundled there, and deps install side-loads it instead (#904; see clauster deps).
  3. Stopping an Interactive Session on Windows is a hard kill: the bridge lives in the keeper's separate ConPTY console, so the graceful CTRL_BREAK can't reach it — the local process is reaped, but the cloud session isn't deregistered (it re-registers on the next launch). Linux/macOS get the orderly double-SIGINT.
  4. Windows dials claustrum's named pipe (discovered via rpc.pipe) rather than the AF_UNIX socket, and clauster spawns the daemon with -listen-pipe plus a -token-file handoff (a numeric token fd isn't usable there). Round-trip validated (#902).
  5. Subscription sign-in (claude auth login, plain pipes) works everywhere; the long-lived setup-token flow runs under a POSIX PTY on Linux/macOS and a ConPTY (pywinpty) on Windows (#905). Because a ConPTY echoes written input back — which a parent can't disable the way POSIX termios does — the operator-pasted code is registered and redacted out of any returned/logged output. Like the Interactive Session, the Windows ConPTY path needs the pty extra (pip install 'clauster[pty]'); without it setup-token fails closed with a clear message (subscription sign-in still works).
  6. Routes exercised on the Windows CI cells + VM.
  7. psutil.cpu_times / memory_info on every platform.
  8. psutil has no per-process io_counters on macOS, so a bridge card's disk_read_bps / disk_write_bps fields are blank there.
  9. An in-process lock serializes clauster's own concurrent writers on every OS; POSIX additionally takes an advisory fcntl.flock, which also guards other clauster processes — the config/CLAUDE.md writers, and since #949 the per-project spawn/stop/forget/adopt sections, so a headless CLI/MCP writer serializes against the live server. Cross-clauster-process serialization therefore needs the POSIX flock; a single clauster process is fully covered everywhere. Neither lock coordinates with the claude CLI (which takes no lock) — that's covered by the atomic os.replace (no torn files) plus the caller's external-edit hash guard.
  10. POSIX sets 0700/0600 mode bits; Windows sets an equivalent owner-only ACL on the state dir via icacls (remove inheritance; grant only the current user + SYSTEM). The Windows ACL is best-effort: the default state_dir (under %USERPROFILE%) already inherits a user + SYSTEM + Administrators ACL, so if icacls can't run (absent, a domain/service account with no resolvable USERNAME, a non-zero exit) clauster logs a loud warning and proceeds on the inherited ACL rather than blocking every state write. Relocate state_dir outside the user profile and you should tighten it yourself.
  11. POSIX fsyncs the parent directory after a rename; Windows can't fsync a directory handle, but NTFS journals the metadata so the rename is recovered on reboot — equivalent durability via a different mechanism.
  12. ops.render_service_unit renders a systemd unit on Linux, a launchd plist on macOS (launchd kind), and a Shawl service-install script on Windows (windows kind). Windows requires Shawl (clauster deps install shawl) before running the generated script. The generated systemd unit sets KillMode=process so live pty bridges survive a systemctl restart (under systemd's default KillMode=control-group they would be reaped — clauster doctor warns when an installed unit is set that way); KillMode is a systemd concept, so this applies on Linux only — launchd and the Windows service have no equivalent.