Muse Code on Windows: PowerShell Setup and Troubleshooting

Install from PowerShell

Open PowerShell in your normal Windows account and use Meta’s official Windows installer. These commands belong in PowerShell, not Command Prompt or a Bash window. Start with the official installer before adding custom wrappers or copying a binary from another machine.

irm https://dev.meta.ai/install.ps1 | iex
muse --version
Get-Command muse -All

The version confirms that Muse starts. Get-Command shows how the shell resolves its name. If installation succeeds but resolution fails, open a new PowerShell window and repeat those two checks. A terminal opened before installation can retain an older PATH. Do not reinstall repeatedly without first checking which command the shell sees.

Check the executable and the working directory

Get-Command muse -All | Select-Object CommandType, Source
Get-Location
Set-Location "C:\src\sample-project"
muse --model muse-spark-1.3 --reasoning-effort medium

Replace the sample path with an existing repository. Quoting it also handles spaces. If more than one Muse command appears, compare their locations and versions before removing anything. A stale wrapper can make two terminals seem to run different releases. The command output is better evidence than the installer’s completion message.

Start the session from the package you actually want to inspect. For a repository containing several applications, name the relevant subdirectory in your request. When Muse cites files, confirm that they come from that checkout rather than a similarly named project elsewhere.

Confirm the intended account without printing secrets

Complete Muse’s browser sign-in flow on Windows. Use /login inside the interactive session if you need to reopen it. Authentication on a Mac does not prove that the Windows host is signed in. Follow the official authentication flow on each machine.

if (Test-Path Env:META_API_KEY) {
  "META_API_KEY is set in this PowerShell process"
} else {
  "META_API_KEY is not set in this PowerShell process"
}

This check reports presence without displaying the key. Use it if a successful browser login still seems to use a different billing route. Inspect where your own shell configuration sets that variable before changing it. Keep credentials out of screenshots, support messages and repository files. See Muse billing and privacy for the distinction between API usage and subscription access.

Ask for commands that work on Windows

Read this project’s README and build configuration.
Identify the test command for Windows PowerShell.
Explain any Bash-only setup steps. Do not modify files.

If the response suggests export or a Unix-only helper, ask for a PowerShell equivalent and compare it with the project’s documented Windows setup. A cross-platform language does not guarantee cross-platform scripts. Build tooling may still depend on Unix paths, shell quoting or an executable that is absent from this host.

For the next task, choose one test that already runs on Windows. Preserve its output before changing code. When a later failure appears, you can distinguish a regression from a pre-existing environment problem. For the full first-session sequence, use the Muse Code getting-started guide.

Know which Windows differences matter

Meta documents a possible UAC prompt when the Windows sandbox first initializes; normal use does not require an administrator terminal. A sandbox error should be investigated as an environment problem before changing permissions. The permissions reference describes the boundary and startup behavior.

Workflow availability also depends on the installed build and rollout. Check the commands exposed by muse --help before following a workflow example. The MCP, skills and workflows guide explains how to separate unavailable features from misconfigured ones.

Windows status in CodeAgentSwarm

When reporting a Windows issue, include the Muse version, Windows architecture, resolved command path and exact error. Describe whether it happened before login, during sandbox setup or after a prompt. That short report helps reproduce the failure without exposing your account or project contents.