claudep

Run Claude Code under separate accounts on one machine, for example a company Enterprise org and a personal Max plan, with fully isolated logins and shared configuration.

Claude Code has no account switching of its own. claudep is one bun script with no dependencies. It manages named profiles and points Claude Code at them through CLAUDE_CONFIG_DIR.

claudep init enterprise --sso --email you@company.com --alias eclaude
eclaude          # Claude Code as the enterprise account
claude           # Claude Code as whatever ~/.claude is logged in as

Not affiliated with Anthropic

claudep is an independent, unofficial tool written by me, Gustavo Bordoni. I have no affiliation with Anthropic. Anthropic has not endorsed, sponsored, reviewed or supported this project in any way. "Claude" and "Claude Code" are Anthropic's names and trademarks and appear here only to describe what the tool works with. Use it at your own risk and under the terms of your own Claude account.

Install

Requires bun and an existing Claude Code install (claude on your PATH).

bun add -g @bordoni/claudep        # or: npm install -g @bordoni/claudep  (bun is still what runs it)
bunx @bordoni/claudep --help       # try it without installing

The package name is @bordoni/claudep and the command it installs is claudep. The same package is mirrored to GitHub Packages, which needs a token with read:packages even for public installs; .ref/releasing.md has the .npmrc lines.

Or clone and symlink it anywhere on your PATH:

git clone git@github.com:bordoni/claudep.git ~/workspace/claudep
ln -s ~/workspace/claudep/claudep.ts ~/.local/bin/claudep

Windows

claudep runs natively on Windows from PowerShell 5.1, PowerShell 7 and Git Bash. Install it with bun, which writes claudep.exe into %USERPROFILE%\.bun\bin:

bun add -g @bordoni/claudep

Two things differ from macOS:

• Profiles share configuration through symlinks, and Windows lets a normal user create symlinks only with Developer Mode on (Settings > For developers > Developer Mode). claudep init stops and says so when it is off. Turn it on, open a new terminal and run the same command again.
• There is no keychain. Claude Code keeps each profile's login in <profile>\.credentials.json, and claudep doctor checks that the file is there.

claudep env prints PowerShell syntax when run from PowerShell and sh syntax from Git Bash, so claudep env work | Invoke-Expression and eval "$(claudep env work)" both work. Git Bash otherwise behaves like bash elsewhere: put eval "$(claudep shell-init bash)" in ~/.bashrc. An npm-installed claude.cmd runs through cmd.exe; the native installer's claude.exe avoids that hop and is what claudep doctor recommends. cmd.exe itself is not supported.

Usage

claudep <name> [claude args…]        run claude with profile <name>
claudep init <name> [options]        create/update a profile and log in
claudep list [--json]                every profile and who it is logged in as
claudep status <name> [--json]       login state for one profile ("default" = ~/.claude)
claudep env <name> [--shell <sh>]    print the CLAUDE_CONFIG_DIR pin for eval, source or Invoke-Expression
claudep alias <name> <command>       write a shim so "<command>" == "claudep <name>"
claudep current [--json|--name]      which profile this shell is on, and why
claudep doctor [name]                verify symlinks, keychain entry, unclassified files, Claude Code version
claudep rm <name> [--keep-login]     log out and delete a profile (base is never touched)
claudep local <name> | --remove      pin the current directory tree to a profile (see below)
claudep resolve [dir]                print the profile pinned for a directory
claudep shell-init [zsh|bash|fish|powershell]   print the hook that applies pins on cd
claudep completion [zsh|bash|fish]   print tab completions for commands, flags and profile names
claudep --version

init options: --sso, --email <addr>, --console, --copy-mcp, --alias <command>, --no-login, --force. Run claudep help for the full text.

Pin a profile to a directory

A file named .claudep at the top of a repo names the profile every hooked shell should use inside that tree, the way .nvmrc names a Node version. Commit it and the whole team inherits the pin.

cd ~/work/acme
claudep local enterprise          # writes ./.claudep containing "enterprise"

Shells apply pins through a hook. Add one line to ~/.zshrc (or ~/.bashrc with bash, including Git Bash on Windows):

eval "$(claudep shell-init zsh)"

In fish the line goes in ~/.config/fish/config.fish, and in PowerShell in $PROFILE:

claudep shell-init fish | source

claudep shell-init powershell | Out-String | Invoke-Expression

From then on, cd into a pinned tree sets CLAUDE_CONFIG_DIR for that profile and cd out of it returns the shell to ~/.claude. The hook is pure shell with no subprocess (builtins in fish, cmdlets only in PowerShell), so it costs nothing at the prompt. The rules:

• The nearest .claudep file upward from the current directory wins. An empty one cancels a parent pin.
• The hook only changes a CLAUDE_CONFIG_DIR it set itself. It tracks that in CLAUDEP_AUTO, so a manual pin from eval "$(claudep env work)" (claudep env work | source in fish, claudep env work | Invoke-Expression in PowerShell) or a plain export stays put until you eval "$(claudep env --unset)".
• A pin that names a profile you have not created prints one warning per directory change and sets nothing.

claudep env prints the syntax of the shell it runs in: it looks at its parent process (fish, pwsh, or an sh-like shell) and falls back to your login shell. Pass --shell sh|fish|powershell when neither answer fits, for example from a script or a Makefile.

claudep current tells you which profile the shell is on and how it got there (hook, manual pin, or nothing). claudep list adds the same line at the bottom.

Tab completion

One line in your rc file completes subcommands, their flags, the values --shell and shell-init take, and your profile names:

eval "$(claudep completion zsh)"      # ~/.zshrc, after compinit
eval "$(claudep completion bash)"     # ~/.bashrc

claudep completion fish | source      # ~/.config/fish/config.fish

Profile names come from a directory listing of ~/.claudep at the moment you press Tab, so the script never runs claudep and a new profile shows up without reloading anything. In zsh the same output also works as a file: claudep completion zsh > ~/.zfunc/_claudep with ~/.zfunc on fpath before compinit, if you prefer autoloading over eval.

Show the profile in your prompt or statusline

In a shell prompt use the variable itself; no command runs:

# zsh or bash: the last path segment is the profile name when a profile is active
PROMPT='${CLAUDE_CONFIG_DIR:+[${CLAUDE_CONFIG_DIR##*/}] }'"$PROMPT"

# fish
function fish_prompt; set -q CLAUDE_CONFIG_DIR; and echo -n "[$(string replace -r '.*/' '' -- $CLAUDE_CONFIG_DIR)] "; ...; end

In scripts, and in the shared statusline-command.sh, claudep current --name prints exactly one word: the profile name, default for ~/.claude, or custom for a CLAUDE_CONFIG_DIR outside the profiles root. The statusline script inherits Claude Code's environment, so it sees the profile the session was started with.

How it works

~/.claude is left exactly as it is and stays the default profile. Each named profile is a thin directory under ~/.claudep/<name>. Inside it, shared configuration is a symlink back into ~/.claude:

Shared (symlinked)	Per profile (real files)
CLAUDE.md and other top-level *.md	.claude.json (login identity, user-scope MCP servers, folder trust)
settings.json, keybindings.json, statusline-command.sh	org-pushed remote-settings.json, policy-limits.json
hooks/, skills/, commands/, agents/, rules/, output-styles/, themes/, workflows/	history.jsonl, todos/, sessions/, teams/, caches, telemetry
plugins/, plans/, projects/ (session transcripts and auto-memory)	credentials

On macOS, credentials never touch the profile directory: Claude Code stores them in the Keychain under Claude Code-credentials-<sha256(CLAUDE_CONFIG_DIR)[0:8]>, so every profile has its own login and refresh token and they cannot overwrite each other. I verified this against Claude Code 2.1.259. An older bug where every config dir shared one Keychain entry no longer applies. On Linux and Windows the login is <profile>/.credentials.json, a real file inside the profile that is never shared.

The shared list is an allowlist, so an account-specific file cannot leak across profiles by accident. claudep doctor reports any base file that is neither shared nor known-private. That is how you notice when a new Claude Code version adds something.

Notes

• The first time you open a repository under a new profile you will re-accept folder trust, and claude.ai connectors need their OAuth redone in that profile.
• Two profiles running at the same time write the same settings.json and plugins/. That is the same situation as two terminals today.
• Background sessions and the daemon are tied to ~/.claude. As of Claude Code 2.1.263, claude daemon install refuses to run with CLAUDE_CONFIG_DIR set, and claude --bg under a profile runs without the daemon.
• Set CLAUDE_PROFILES_DIR to move the profiles root. Keep it out of iCloud or Dropbox; .claude.json is rewritten constantly and sync tools create conflict copies.
• Never put CLAUDE_CONFIG_DIR in a settings.json env block. Claude Code detects that mismatch and disables features; claudep doctor reports it.
• ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN and CLAUDE_CODE_OAUTH_TOKEN override the login in any config dir, always in -p mode. claudep <name> and claudep doctor warn when one is set and leave it alone.
• claudep doctor fails on macOS when Claude Code is older than 2.1.144, the first build whose Keychain item is namespaced per config dir.

Releases

Versions, dates and changes are in CHANGELOG.md. Releases are tagged with the bare version, 0.1.0, and published to npm with provenance by GitHub Actions.

Development

bun install        # dev tooling only; nothing ships with the tool
bun run check      # typecheck, lint, tests. Same as CI

To work on the script, start with AGENTS.md. The detail lives in .ref/.

License

MIT