Repository navigation
Description
Product or interface
Desktop app
Version
- Desktop app (client): 3.1.16 (macOS)
- Remote CLI
@minimax-ai/code: 0.7.7 (Windows x64, installed manually viainstall.ps1) - Remote CLI installed via the official Windows installer is reported as 0.7.7
Platform
Multiple platforms
OS version and architecture
- Remote host: Windows 11,
Microsoft Windows NT 10.0.26100.0, x64 - Client: macOS 15.7.7, arm64
Issue area
Startup / install / update
Steps to reproduce
- Install MiniMax Code Desktop on a macOS machine.
- Open Settings → Connections → SSH → Add, and pick a host entry that points at a Windows machine over SSH (host
192.168.5.x, port 22, identity file~/.ssh/<key>). The host entry is read from~/.ssh/config. - The connection is saved and enabled, but the host stays in a failed state with: "Remote MCode installation failed. Please check the developer machine network and retry."
- Click Install MCode on that host entry. The remote installer fails again, and the remote machine gets a new log file under
~/.minimax-code-install-logs/install-*.log. - Install the CLI on the Windows host manually instead:
This succeeds and
irm https://filecdn.minimax.chat/public/install.ps1 | iex
mcode --versionreports0.7.7. - With the CLI present, the Desktop app still cannot connect. The app-server it needs on the remote host refuses to start on Windows (see the error summary below).
Expected and actual behavior
Expected: A Windows machine reachable over SSH can be registered as an SSH remote host in Settings → Connections → SSH, and the Desktop app can drive the MiniMax Code runtime installed there — the same way a Linux host works today.
Actual: Two independent Windows-specific failures block this:
-
The Desktop remote installer runs the Unix installer on Windows. It executes
install.shon the Windows host, and that script rejects the platform:==> [1/5] Checking platform and install location error: Windows is not supported by install.sh; run the install.ps1 command in PowerShell.The official
install.ps1works correctly when run manually in PowerShell on the same host. -
The remote app-server refuses to start on a Windows host. Even after the CLI is installed correctly, the app-server the Desktop app needs to connect to is rejected:
mcode app-server daemon bootstrap --auth-mode remote-oauth {"action":"bootstrap","error":{"code":"appServerDaemonCommandFailed", "message":"The app-server daemon requires a Unix-like remote host.", "recommendedAction":"inspectRemoteLogs"}}mcode app-server --listen unix://<path> --foreground The app-server Unix transport requires a Unix-like remote host.Both the daemon path and the direct
--listen unix://path are refused, so this is not a configuration problem — the Windows host cannot host the app-server at all.
Redacted error summary
# Desktop side (macOS client), Settings -> Connections -> SSH
[info] [SshRuntime] connection failed: The remote CLI path could not be discovered.
[info] [SshRecovery] {"hostId":"<windows-host>","state":"action_required","phase":"stopped",
"code":"remote_mcode_not_found","disposition":"action_required"}
# Windows remote host, Desktop "Install MCode" action
[mcode-install] operationId=manual-<id> stage=1
==> [1/5] Checking platform and install location
error: Windows is not supported by install.sh; run the install.ps1 command in PowerShell.
[mcode-install] operationId=manual-<id> stage=1 exitCode=1
# Windows remote host, manual CLI install -> succeeds
> irm https://filecdn.minimax.chat/public/install.ps1 | iex
... Installation complete: @minimax-ai/code@0.7.7 (fresh)
> mcode --version
0.7.7
# Windows remote host, app-server bootstrap (the blocker)
> mcode app-server daemon bootstrap --auth-mode remote-oauth
{"action":"bootstrap","error":{"code":"appServerDaemonCommandFailed",
"message":"The app-server daemon requires a Unix-like remote host.",
"recommendedAction":"inspectRemoteLogs"}}
> mcode app-server --listen unix://<path> --foreground
The app-server Unix transport requires a Unix-like remote host.
Screenshots
The Settings → Connections → SSH editor for the Windows host is filled in correctly (host 192.168.5.x, port 22, identity file ~/.ssh/<key>), and the host entry shows the "Remote MCode installation failed" state. Screenshot can be provided on request.
Additional notes
- The SSH connection configuration itself is correct, and plain SSH to the same host works fine (
mcode --versionreturns0.7.7over a normal SSH session). - The CLI on Windows is fully functional for headless use:
mcode exec "..."runs successfully on the Windows host and returns a proper result. Only the Desktop-to-remote app-server transport is affected. - Related existing issues: [Bug]: SSH remote app-server never becomes ready on slow hosts, and a stale socket breaks the connection permanently #545 (slow hosts / stale socket) and Reduce remote app-server resident footprint - make the updater opt-out and publish minimum hardware requirements #546 (remote app-server footprint). Both discuss Linux remotes only; neither covers a Windows remote host.
- Searched this repository for existing reports: there is no issue or feature request about Windows as an SSH remote host — all Windows reports are about Windows as a client.
Before submitting
- I have searched existing issues.
- I have included my version and removed sensitive information.
Question for the maintainers
Is Windows supported (or planned) as an SSH remote host for the Desktop app? If it is not, could the Desktop app either:
- detect a Windows remote host and run
install.ps1instead ofinstall.sh, and - surface a clear "Windows remotes are not supported yet" message instead of the current generic installation failure?
Documenting the current limitation in the Connections settings page would also help — currently the SSH remote feature has no documentation page, and the only platform guidance I could find is in the changelog.
Thanks for taking a look.
Activity
Ralle1976 commentedon Nov 20, 2026
Reachability balance first (the label is
desktop, so this comes before the measurements)I looked for both halves of your report in this repository's published source and found neither one there. That is why I am commenting here with measurements rather than a patch.
1. The Desktop half is out of scope by the repository's own publication rule.
docs/publication-authorization.md:15— "The published source does not include the Desktop application's source, internal Git history, private review material, account data, logs, sessions, credentials, or later unreviewed additions."docs/tui-capabilities.md:212repeats it for the runtime side — "Desktop HTTP services and internal generated IDL remain outside the public repository." SoSettings → Connections → SSH, the~/.ssh/confighost entries, the "Install MCode" button and theinstall.sh-vs-install.ps1choice cannot be fixed from this tree — that is yourSshRuntime/SshRecoverylog side.2. The CLI half is also not in the published source, although it is in the shipped npm package.
This is the part worth stating precisely, because it is easy to assume the opposite.
3c21163git grep -c -i "app-server"packages/local-runtime/src/api/host-route-contexts.ts:175git grep -n "Unix transport",git grep -n "unsupported_platform"git log --all -S "requires a Unix-like remote host"packages/tui/src/cli/program.ts:95-359init,exec,acp,login,logout,update,telemetry,status,preview,provider,plugin, … — noapp-servercommandscripts/lib/package-exports.mjs:18-45fromrelease/extraction.json→packageRoots)app-server/ssh/daemon/remotescripts/build.mjs:80-81cli: "packages/tui/src/index.ts"— the published CLI is built frompackages/tuionlyThe npm package you installed does contain the command (
mcode app-server --help→ "Run the MCode SSH app-server"). The source of that command is simply not part of this repository's public projection, so no PR against this repo can move the Windows gate you hit.docs/open-source-status.md:3scopes the tree to "the terminal TUI, headless CLI, ACP implementation, and the public distribution tooling around them" — the app-server is none of those.So the honest answer to "is this an OSS-repo bug?" is: no, not one that this repository can fix. What this repository can contribute is the measurement below, because the gate you hit is a transport choice, not a broken Windows runtime — and that is the one dimension your report does not contain.
Contribution: the Windows refusal is bound to the transport, not to the app-server
Your two quoted failures are real and I reproduced both byte-for-byte on a Windows host. But they are two different gates, and neither of them is the app-server itself.
Versionsanker
mcode --version→0.7.8(global npm install)npm view @minimax-ai/code version→0.7.8(latest), installed == publishedmain=3c21163(release/extraction.json→sourceRevisionbaseline9b9885e; rootpackage.jsonversion0.7.2)v24.18.010.0.26200, x64%TEMP%lab with--data-dirpointed at it;~/.minimaxwas not modifiedssh.exe, client only — nosshdon this host, so no end-to-end SSH handshake was possible)Repro command (Windows, isolated, one block)
Measured matrix on Windows (CLI 0.7.8):
app-server --listen unix://<path> --foregroundThe app-server Unix transport requires a Unix-like remote host.app-server daemon bootstrap{"action":"bootstrap","error":{"code":"appServerDaemonCommandFailed","message":"The app-server daemon requires a Unix-like remote host.","recommendedAction":"inspectRemoteLogs"}}, exit 1app-server daemon startapp-server daemon statusapp-server daemon version"version":"0.7.8"app-server proxyconnect ENOENT <data-dir>\app-server-control\app-server.sockbecause no server can be listening thereapp-server --listen stdio:// --foregroundapp-server --stdio --foregroundThe first two rows are identical to your quoted output, so the reproduction is confirmed; the interesting part is the last two.
The
stdio://transport completes the full app-server handshake on Windows. I sent a realinitializerequest over stdio and got a realresultback (no error, process stayed alive):{"id":1,"result":{"protocolVersion":1, "serverInfo":{"name":"mcode-app-server","title":"MCode SSH App Server","version":"0.7.8"}, "environment":{"platformFamily":"windows","platformOs":"win32","arch":"x64","userName":"<redacted>","homeDir":"<redacted>","dataDir":"<redacted>"}, "hostIdentity":{"authMode":"remote-oauth","protocolVersion":1,"version":"0.7.8","region":"en","buildEnv":"prod"}, "capabilities":{"methods":["account/read","account/login/start","account/login/cancel","account/logout","server/prepareRestart","desktopService/transport"], "transports":["stdio","unix-websocket"], "desktopService":{"contract":"DesktopService","transportMethod":"desktopService/transport","transportProtocolVersion":2,"binaryPayloadEncoding":"base64",...}, "features":["project.icon.update","computer.use"], "reconnect":{"history":true}, "auth":{"deviceAuthorization":true,"credentialTransfer":false,"mode":"remote-oauth"}}}}Three things in that payload are directly relevant to your question:
environment.platformFamily: "windows"/platformOs: "win32"— the runtime is Windows-aware; it reports its own platform instead of refusing to start.capabilities.transports: ["stdio","unix-websocket"]— the server advertisesunix-websocketon this very win32 host, while the same binary refuses--listen unix://there. The advertisement and the gate disagree.capabilities.featurescontainscomputer.use— the exact feature the CLI changelog attributes to a Windows remote (see below).So the Windows-specific mechanism is a transport/daemon gate, and it is narrower than "the app-server does not run on Windows". Read from the published bundle (
node_modules/@minimax-ai/code/chunks/run-app-server-command-EWT4AIHT.js, minified, single-line chunk — hence function names instead of source lines):Ci(listenUri)"stdio"/"stdio://"→{kind:"stdio"};"unix://…"→{kind:"unix",socketPath}; anything else →Unsupported app-server listen URI: … Use stdio:// or unix:///path.Ri(...)(server bootstrap)if (s.kind === "unix" && platform === "win32") throw new Error("The app-server Unix transport requires a Unix-like remote host.")— the refusal is conditioned on the transport kindstart()if ((options.platform ?? process.platform) === "win32") throw new C("unsupported_platform", "The app-server daemon requires a Unix-like remote host.")— unconditional for the daemondn(...)if (platform === "win32") throw new Error("The app-server daemon requires a Unix-like remote host.")n = process.getuid?.()→n === undefinedon Windows →App-server control directory must be a real directory owned by the current user: …; and(mode & 0o777) !== 0o700→App-server control directory requires effective 0700 permissions; use a private directory on a filesystem supporting Unix permissions: …proxymoderunProxy({ socketPath, input: stdin, output: stdout })— connects to a Unix socket pathWhy the Unix transport cannot simply be enabled on Windows — measured, not assumed (Node
v24.18.0, this host):net.createServer().listen("<fs path>")— 4 path variants (relative, drive-letter, backslash, short%TEMP%path)EACCES: permission deniedon every variantnet.createServer().listen("\\\\.\\pipe\\<name>")process.getuidon Windowsundefined0o666, not0o700So the Windows-native IPC primitive for this design is a named pipe, not a Unix socket — and the daemon's control-directory ownership/permission preconditions are Unix-shaped by construction, independent of the socket itself. That is a design consequence, not a small portability bug, and I am not proposing a change to it.
A documentation inconsistency worth closing on your side of the report. (Note that these entries live in the
CHANGELOG.mdof the published npm package; the published source tree at3c21163contains no markdown that mentions the app-server at all.) The shipped CLI changelog already describes Windows as a valid app-server host:CHANGELOG.mdof the published package (v0.7.8,node_modules/@minimax-ai/code/CHANGELOG.md:82), entry 0.7.6 (2026-11-05): "When Desktop opens a project over SSH,mcode app-serveron the remote macOS or Windows machine provides Computer Use there; permissions are granted on that machine with/computer-use permissions."CHANGELOG.md:250, entry 0.6.12 (2026-10-15): "A standalone App Server supports SSH connections to a remote Runtime, including connection recovery, authentication state, and version updates."and the handshake above confirms that a Windows remote can carry the
computer.usefeature over stdio. Meanwhile the same binary refuses the Unix transport and the daemon onwin32. I cannot tell from here which of the two is the intended product boundary — that is a question for the maintainers, and it is the same question your report asks.