claude-profile

Manage multiple Claude Code configuration profiles. Switch between work and personal accounts, different MCP server setups, or separate settings without logging in and out.

Each profile is a complete, isolated Claude Code configuration directory (settings, credentials, MCP servers, CLAUDE.md, history -- everything). Once configured, claude automatically uses your active profile -- no special launch command needed.

Install

Linux / macOS / WSL / Git Bash (MSYS2):

curl -fsSL https://raw.githubusercontent.com/pegasusheavy/claude-code-profiles/main/install.sh | sh

Windows (PowerShell):

irm https://raw.githubusercontent.com/pegasusheavy/claude-code-profiles/main/install.ps1 | iex

The installer downloads the appropriate scripts and configures your shell. Restart your shell (or open a new terminal) after installing.

Quick Start

# Create profiles
claude-profile create work
claude-profile create personal

# Set a default
claude-profile default work

# Just use claude — it automatically uses your default profile
claude
claude --resume
claude -p "explain this code"

Commands

Command	Description
claude-profile	Show current profile status
claude-profile use <name>	Switch to a profile for this session
claude-profile create <name>	Create a new profile
claude-profile list	List all profiles
claude-profile default [name]	Get or set the default profile
claude-profile local [name]	Show, set, or --remove this directory's .claude-profile
claude-profile auto [on|off|status]	Control directory-local auto-switching (not on cmd.exe)
claude-profile skills	List pool skills and their status for the profile
claude-profile skills register <name> <path>	Add a skill directory to the shared pool
claude-profile skills unregister [--force] <name>	Remove a skill from the pool
claude-profile skills add <skill> [profile]	Add a pool skill to a profile
claude-profile skills remove <skill> [profile]	Remove a pool skill from a profile
claude-profile skills set <s1,s2,...> [profile]	Replace a profile's skill selection
claude-profile skills reset [profile]	Back to the default (all pool skills)
claude-profile skills sync [profile|--all]	Re-materialize skill links
claude-profile delete <name>	Delete a profile (with confirmation)
claude-profile which [name]	Show the config directory path
claude-profile version	Show the installed version
claude-profile update [--force]	Update to the latest release
claude-profile help	Show help

How It Works

Claude Code supports a CLAUDE_CONFIG_DIR environment variable that redirects where it stores configuration and data. claude-profile provides a claude() shell function that wraps the real claude binary:

1. On each directory change, the nearest .claude-profile file (if any) sets CLAUDE_CONFIG_DIR — see Per-Directory Profiles.
2. Before each invocation, the wrapper checks if a default profile exists and auto-sets CLAUDE_CONFIG_DIR.
3. If CLAUDE_CONFIG_DIR is already set (e.g., via claude-profile use), it is used as-is.
4. The real claude binary is then called with all your arguments.

This means you never need to think about profiles during normal use -- just run claude as you always have.

Session Override

To temporarily use a different profile in the current shell session:

# Temporarily use a different profile
claude-profile use personal
claude                          # uses "personal" for this shell session

The override lasts until you close the shell or run claude-profile use again.

Per-Directory Profiles

A directory can pin itself (and everything under it) to a profile by holding a .claude-profile file whose first non-empty, non-comment line is a profile name:

cd ~/work/acme
claude-profile local work        # writes ./.claude-profile containing "work"

Your shell now switches to work whenever you cd into that tree, and back to the default when you leave:

$ cd ~/work/acme/api
claude-profile: switched to profile 'work' (from /home/you/work/acme/.claude-profile)
$ cd ~
claude-profile: directory profile cleared; using the default profile

The file is plain text — you can commit it to a repo, and blank lines and # comments are ignored:

# every checkout of this repo uses the client's isolated profile
client-acme

Precedence and control:

• An explicit claude-profile use <name> pins the session: it wins over any .claude-profile until you run claude-profile auto on, which clears the pin and re-resolves the current directory.
• claude-profile auto off disables auto-switching for the session; claude-profile auto status shows what's in effect.
• CLAUDE_PROFILE_NO_AUTO_SWITCH=1 disables it entirely, CLAUDE_PROFILE_AUTO_QUIET=1 switches silently.
• A .claude-profile naming a profile that doesn't exist (or an invalid name) is reported on stderr once per directory entry and otherwise ignored.
• claude-profile local --remove deletes the .claude-profile in the current directory.

Hooking into directory changes is per-shell: zsh uses chpwd, bash uses PROMPT_COMMAND, and other POSIX shells get a cd wrapper. PowerShell 6+ uses LocationChangedAction; Windows PowerShell 5.1 hooks the prompt function.

cmd.exe is different. It has no cd hook and no transparent claude wrapper, so it can't switch automatically. Instead, a bare call claude-profile.cmd prefers the nearest .claude-profile over the default profile, and claude-profile local <name> writes the file:

claude-profile local work
call claude-profile
claude

Per-Profile Skills

Skills usually have to be curated by hand per config directory. claude-profile can instead manage them from a shared pool at <data-dir>/skills/ (next to your profile directories), with each profile selecting the subset it wants — so a backend profile only loads backend skills, keeping context lean:

# Register skills into the pool (each path must contain a SKILL.md)
claude-profile skills register api-design ~/skills/api-design
claude-profile skills register ui-design  ~/skills/ui-design

# Pick subsets per profile
claude-profile skills set api-design backend
claude-profile skills set ui-design frontend

# Or tweak incrementally
claude-profile skills add ui-design backend
claude-profile skills remove ui-design backend

How it works:

• Registering creates a symlink (a directory junction on Windows — no admin rights needed) in the pool; the pool entry name is what profiles refer to.
• A profile's selection lives in <profile>/skills.conf — plain text, one skill name per line, # comments allowed. No skills.conf means the profile gets every pool skill, so existing profiles keep working unchanged. An empty file means none.
• Syncing materializes the selection as links in <profile>/skills/. The tool only ever creates or removes links that point into the pool — hand-made symlinks, real directories, and files in <profile>/skills/ are never touched (a name collision is reported and the unmanaged entry wins).
• Links update only when you run a skills command (or at claude-profile create) — switching profiles does no filesystem work. After registering or unregistering pool skills, run claude-profile skills sync --all to propagate.
• The profile name skills is reserved for the pool directory.

Filtered profiles are annotated in claude-profile list (e.g. work [skills: 3/12]) and in the bare claude-profile status output (Skills: 3 of 12 pool skills (filtered)). (cmd.exe has no bare status command, so there the annotation appears in list only.)

Profile Storage

Profiles are stored in platform-appropriate locations:

Platform	Location
Linux	$XDG_DATA_HOME/claude-profiles/ (default: ~/.local/share/claude-profiles/)
macOS	$XDG_DATA_HOME/claude-profiles/ (default: ~/.local/share/claude-profiles/)
Windows	%LOCALAPPDATA%\claude-profiles\
Git Bash / MSYS2	%LOCALAPPDATA%\claude-profiles\ (shared with cmd/PowerShell)

Each profile directory is a complete Claude Code config directory. After creating a profile and launching Claude with it, Claude will populate it with settings.json, .credentials.json, and everything else it needs.

Profile Names

Profile names can contain letters, digits, hyphens, and underscores. Examples: work, personal, client-acme, side_project.

Updating

claude-profile checks for new releases at most once every 24 hours, as a side effect of running claude (or, on cmd.exe, any claude-profile command). If a newer version is available, you'll see a one-time notice:

A new claude-profile version is available (v1.0.0 -> v1.1.0). Run 'claude-profile update' to upgrade.

Run the update:

claude-profile update

This downloads the new script and the shared VERSION file from the matching GitHub Release (not the main branch), verifies their SHA-256 checksums, and only replaces your installed files once both checks pass. It refuses to downgrade unless you pass --force. After updating, restart your shell (or source ~/.bashrc / . $PROFILE) to pick up the new version — cmd.exe re-reads the file on every call, so no restart is needed there.

To disable the passive check entirely, set:

export CLAUDE_PROFILE_NO_UPDATE_CHECK=1

If the update-check cache itself can't be written (permissions, a full disk), you'll see a similar one-time-per-day warning instead:

claude-profile: warning: could not write update-check cache in ~/.local/share/claude-profile -- update notifications may not work until this is fixed

This doesn't block claude from running — it's just letting you know update notifications may be unreliable until the underlying issue is fixed.

Platform Support

Script	Platform	Shell
claude-profile.sh	Linux, macOS, WSL, Git Bash / MSYS2	bash, zsh (sourced)
claude-profile-init.ps1	Windows, Linux, macOS	PowerShell 5.1+ / pwsh 6+ (dot-sourced)
claude-profile.cmd	Windows	cmd.exe (use with call prefix)

Git Bash / MSYS2 Support

On Git Bash and other MSYS2-based shells on Windows, claude-profile.sh automatically detects the environment and:

• Stores profiles at %LOCALAPPDATA%\claude-profiles\ (shared with cmd.exe and PowerShell implementations)
• Converts Unix-style paths to Windows-native paths via cygpath -w before passing them to the native claude.exe binary
• Uses the MSYSTEM environment variable for detection (MINGW64, MINGW32, MSYS, etc.)

This means Git Bash users share the same profile data with cmd.exe and PowerShell on the same machine — no duplication or conflicts.

Manual Install

If you prefer not to use the install scripts:

Linux / macOS:

# Download
mkdir -p "${XDG_DATA_HOME:-$HOME/.local/share}/claude-profile"
curl -fsSL https://raw.githubusercontent.com/pegasusheavy/claude-code-profiles/main/claude-profile.sh \
  -o "${XDG_DATA_HOME:-$HOME/.local/share}/claude-profile/claude-profile.sh"

# Add to shell profile (.bashrc or .zshrc)
echo '. "${XDG_DATA_HOME:-$HOME/.local/share}/claude-profile/claude-profile.sh"' >> ~/.bashrc

Windows (PowerShell):

$dir = "$env:LOCALAPPDATA\claude-profile"
New-Item -ItemType Directory -Force -Path $dir | Out-Null
Invoke-WebRequest -Uri "https://raw.githubusercontent.com/pegasusheavy/claude-code-profiles/main/claude-profile-init.ps1" -OutFile "$dir\claude-profile-init.ps1"
Invoke-WebRequest -Uri "https://raw.githubusercontent.com/pegasusheavy/claude-code-profiles/main/claude-profile.cmd" -OutFile "$dir\claude-profile.cmd"
# Add to PowerShell profile
Add-Content -Path $PROFILE -Value ". '$dir\claude-profile-init.ps1'"
# Add to PATH for cmd.exe
$path = [Environment]::GetEnvironmentVariable('Path', 'User')
if ($path -notlike "*$dir*") { [Environment]::SetEnvironmentVariable('Path', "$path;$dir", 'User') }

License

MIT