One Brain, Two Wallets: Two Claude Code Accounts on One Machine

I hit my Claude Code usage limit mid-week. I have a second subscription, but switching accounts the naive way — /logout

, /login

, repeat — is miserable, and it turns out "switching accounts" in Claude Code silently means switching everything: skills, MCP servers, settings, hooks, memory, session history. All of it.

What I actually wanted: one brain, two wallets. Same skills, same MCP servers, same muscle memory — just a different subscription getting billed.

Here's the setup that got me there. ~20 minutes, fully reversible.

Claude Code keeps all its state in a config directory — ~/.claude

by default — and honors a CLAUDE_CONFIG_DIR

env var to point elsewhere. That's the whole trick:

alias claude-work='CLAUDE_CONFIG_DIR=~/.claude claude'
alias claude-personal='CLAUDE_CONFIG_DIR=~/.claude-personal claude'
mkdir ~/.claude-personal
source ~/.zshrc
claude-personal   # prompts OAuth login for the second account — one time

Credentials land in whichever dir is active. Bare claude

still defaults to ~/.claude

, so nothing about your existing workflow changes.

.claude.json

moves when you set the env var Claude Code's user-level config file (~/.claude.json

— MCP servers, folder trust, onboarding state) lives next to ~/.claude

, not inside it. But when CLAUDE_CONFIG_DIR

is set, it's read from inside the config dir instead.

So claude-work

— even though it points at the same ~/.claude

you've always used — creates a fresh, empty ~/.claude/.claude.json

and suddenly has zero MCP servers. Fix: symlink it to the master, since it's the same account:

ln -s ~/.claude.json ~/.claude/.claude.json

Now bare claude

and claude-work

are byte-for-byte identical.

The second account starts as a blank slate. Since only billing should differ, symlink everything shareable from the main dir:

cd ~/.claude-personal
ln -s ~/.claude/skills skills
ln -s ~/.claude/CLAUDE.md CLAUDE.md
ln -s ~/.claude/settings.json settings.json        # hooks, permissions, env
ln -s ~/.claude/settings.local.json settings.local.json
ln -s ~/.claude/plugins plugins

One source of truth: add a skill or tweak a hook once, both accounts see it.

MCP servers need more care. They live in .claude.json

, which also holds the account identity (oauthAccount

) — so you can't symlink the whole file across accounts. Merge just the mcpServers

block:

jq --slurpfile mcp <(jq '.mcpServers' ~/.claude.json) \
   '.mcpServers = $mcp[0]' ~/.claude-personal/.claude.json \
   > /tmp/merged.json && mv /tmp/merged.json ~/.claude-personal/.claude.json

--resume

doesn't know about accounts Session transcripts live at <config-dir>/projects/<project>/<session-id>.jsonl

. Resume a personal session under the work account and Claude just says the session doesn't exist. Worse, anything that launches bare claude

for you — crash-restore scripts, launchd jobs — always resumes against the default dir.

The fix is a tiny shim earlier in $PATH

than the real binary. Every session ID exists in exactly one account's tree, so ownership is the routing signal:

#!/bin/bash
REAL=/opt/homebrew/bin/claude   # wherever `command -v claude` pointed before

if [[ -z "$CLAUDE_CONFIG_DIR" ]]; then
  id="" prev=""
  for arg in "$@"; do
    if [[ $prev == "--resume" || $prev == "-r" ]]; then id=$arg; break; fi
    if [[ $arg == --resume=* ]]; then id=${arg#--resume=}; break; fi
    prev=$arg
  done
  if [[ -n $id ]] && compgen -G "$HOME/.claude-personal/projects/*/$id.jsonl" >/dev/null; then
    export CLAUDE_CONFIG_DIR="$HOME/.claude-personal"
  fi
fi
exec "$REAL" "$@"

Now claude --resume <id>

just works, from anywhere, for either account. Explicit CLAUDE_CONFIG_DIR

(your aliases) passes through untouched.

The whole reason I did this was running out of credits mid-conversation. Since a session is just a transcript file, "continue this work conversation on my personal account" is a copy + resume:

#!/bin/bash
set -euo pipefail
id=${1:?usage: claude-pickup <session-id>}
src=$(ls ~/.claude/projects/*/"$id".jsonl 2>/dev/null | head -1)
[[ -n $src ]] || { echo "session $id not found" >&2; exit 1; }
proj=$(basename "$(dirname "$src")")
mkdir -p ~/.claude-personal/projects/"$proj"
cp -n "$src" ~/.claude-personal/projects/"$proj"/
CLAUDE_CONFIG_DIR=~/.claude-personal exec claude --resume "$id"

The work account keeps its copy as history; new turns land only in the personal copy. And since the shim checks the personal tree first, future bare claude --resume

of that ID follows the conversation to its new wallet automatically.

/mcp

once in the new account to re-auth. CLI-based MCP servers that use machine credentials work immediately.

claude              # work (default, unchanged)
claude-personal     # second subscription, same brain
claude --resume X   # figures out the account itself
claude-pickup X     # mid-conversation wallet swap

Same skills, same MCP servers, same hooks, same memory. The only thing that changes is who gets the bill.

source & further reading

dev.to — original article My OpenClaw Agent Had a Panic Attack at 3 AM. Its Model Fallback Chain Saved It Building a Diversifi Memory Agent on Qwen Cloud Is China's Open-Weights AI Strategy Actually Winning?