Using multiple Claude Code accounts¶

agb can index and launch sessions from multiple Claude Code subscriptions. Each account keeps separate authentication, settings, and session storage while all accounts can work in the same project directories.

Prerequisites¶

• Two or more Claude subscriptions, each using a different Claude account
• Claude Code installed and available as claude
• agb installed from this repository

Install agb locally when needed:

make install

How isolation works¶

Claude Code normally stores its first account in its default home:

• Session data: ~/.claude/projects
• Main metadata: ~/.claude.json

Additional accounts use CLAUDE_CONFIG_DIR. For example, CLAUDE_CONFIG_DIR=~/.claude-cc2 places the second account's credentials, settings, and sessions below ~/.claude-cc2.

Do not set CLAUDE_CONFIG_DIR=~/.claude for the default account. The default layout is special because ~/.claude.json sits outside ~/.claude. Launch the default account without CLAUDE_CONFIG_DIR instead.

1. Create account commands¶

Zsh or Bash¶

Add these functions to ~/.zshrc or ~/.bashrc:

cc1() {
  command claude "$@"
}

cc2() {
  CLAUDE_CONFIG_DIR="$HOME/.claude-cc2" command claude "$@"
}

Reload the shell:

source ~/.zshrc

Use source ~/.bashrc when using Bash.

PowerShell¶

Add these functions to the PowerShell profile shown by $PROFILE:

function cc1 {
    claude @args
}

function cc2 {
    $previous = $env:CLAUDE_CONFIG_DIR
    try {
        $env:CLAUDE_CONFIG_DIR = Join-Path $HOME ".claude-cc2"
        claude @args
    }
    finally {
        if ($null -eq $previous) {
            Remove-Item Env:\CLAUDE_CONFIG_DIR -ErrorAction SilentlyContinue
        }
        else {
            $env:CLAUDE_CONFIG_DIR = $previous
        }
    }
}

Reload the profile:

. $PROFILE

2. Authenticate both subscriptions¶

Run each login separately and choose the corresponding Claude account in the browser:

cc1 auth login
cc2 auth login

Verify that both homes are authenticated:

cc1 auth status
cc2 auth status

Each command should report "loggedIn": true and the expected account. The two config homes do not share credentials.

3. Configure agb¶

Create ~/.config/agentbridge/config.json:

{
  "skipPermissions": false,
  "claudeAccounts": [
    {
      "name": "cc1"
    },
    {
      "name": "cc2",
      "configDir": "~/.claude-cc2"
    }
  ],
  "defaultClaudeAccount": "cc1"
}

The account without configDir is Claude Code's default home. Account names must be unique. They become picker badges such as [CC1] and [CC2].

defaultClaudeAccount is the account Claude routes land on when a route reaches Claude without naming one, and it leads the Claude group when the target is cycled. It does not change which account owns existing sessions.

4. Use the picker¶

Start normally:

agb

Picker controls:

Badges preview the selected route. Until the target is cycled it follows the selection, so every row shows a single label and Enter resumes it in its own account. With CC2 targeted, a CC1 session is shown as [CC1→CC2]; a session already owned by CC2 remains [CC2]. Cycling back to CC1 reverses cross-account routes to [CC2→CC1].

Set the initial target from the command line:

agb --claude-account cc2    # or --target cc2

Continue an old CC1 session with CC2¶

1. Run agb --claude-account cc2.
2. Select the old [CC1] session.
3. Confirm the picker says Target: CC2 and the badge reads [CC1→CC2].
4. Press Enter.

agb reads the CC1 transcript and starts a fresh CC2 session with that context. The original CC1 session remains unchanged. Once Claude Code persists the new session, it appears in agb with a [CC2] badge.

Resume and fork behavior¶

Claude session IDs belong to the config home that created them. agb therefore uses two different paths:

• Same account: launch claude --resume with the owning account's environment.
• Different account: build a fresh continuation prompt from the transcript and launch Claude with the target account's environment.

The same two paths apply to a route that crosses tools, with the target tool's own resume command and environment.

A cross-account fork carries conversation text, not hidden model state or the source account's credentials. Both sessions still use the same working directory, so they see the same project files and Git worktree.

Add another account¶

Adding an account later (here a third, cc3) is the same few steps whichever number it is. Do them in this order; every step can be re-run safely.

1. Give it a wrapper and log in¶

cc3() {
  CLAUDE_CONFIG_DIR="$HOME/.claude-cc3" command claude "$@"
}

Reload the shell, then authenticate with the new subscription's Claude account and check it is the one you meant:

cc3 auth login
cc3 auth status

Settings are per account too, so copy the ones you rely on from another account's settings.json (for example "theme": "auto") into ~/.claude-cc3/settings.json.

2. Tell agb about it¶

Add the account to claudeAccounts in ~/.config/agentbridge/config.json:

{
  "name": "cc3",
  "configDir": "~/.claude-cc3"
}

That is all the picker needs: agb --print lists its sessions as [CC3], and Ctrl+T and agb --target cc3 reach it. Pick a name not used by any other tool's account; the names share one list of targets.

3. Give it the shared setup¶

If you keep skills, MCP servers or memory in a setup profile, add the account to that profile's targets, and to memory.share if the accounts share memory:

"targets": ["oc", "cc1", "cc2", "cc3", "cx"],
"memory": { "share": ["cc1", "cc2", "cc3"] }

Keep the store (the first share entry) where it is; a new account joins it. Preview, then apply:

agb plan     # cc3 rows: skills and MCPs to add, memory folders to link
agb sync
agb plan     # every cc3 row now reads "present"

A brand-new account has no memory of its own, so it only gets links: it starts with everything the other accounts already learned. An account that already has memory is merged into the store first (see setup.md).

4. Re-install per-account integrations¶

Anything that installs itself into a Claude config dir does so for one account at a time: hooks in settings.json, status lines, plugins, and tools that read transcripts from projects/. Install them again with the new account's CLAUDE_CONFIG_DIR, for example a terminal multiplexer's session hook:

CLAUDE_CONFIG_DIR=~/.claude-cc3 herdr integration install claude

A transcript viewer that only knows ~/.claude/projects needs ~/.claude-cc3/projects added to its list of roots.

5. Check¶

cc3 auth status                  # logged in, the right account
agb --print --target cc3         # agb accepts the target
agb plan | grep cc3              # nothing left to add or link

The new account shows sessions in the picker once it has run one conversation; until then only its target exists.

Troubleshooting¶

No CC2 sessions appear¶

An authenticated account has no session files until it starts a conversation. Run cc2 once or fork a session into CC2 with Enter on a [CC1→CC2] row.

The wrong subscription opens¶

Check both identities with cc1 auth status and cc2 auth status. Confirm the second configDir matches the value used by the cc2 wrapper.

Old sessions disappear¶

Keep the first account entry without configDir. This lets agb continue scanning ~/.claude/projects, where existing default-account sessions live.

A session resumes in the wrong account¶

Do not move session JSONL files between config homes. Keep account definitions stable after sessions exist; agb records ownership from the directory it scanned.

Protect credentials¶

Do not copy or commit .credentials.json from any Claude config directory. The agb config contains directory locations only and does not need tokens or passwords.