limitbreak

   ★ L I M I T   B R E A K ★
   ADA    [██████████] LIMIT   → spent
   GRACE  [░░░░░░░░░░] ready   → takes the hit

Use several Claude accounts from one Claude Code install, and switch between them automatically when an account's usage limit is hit, without restarting the session or losing its history.

A small local proxy (a single Go binary, standard library only, a few MB of RAM) sits between claude and Anthropic's API. It sends each request on the active account's own login. When Anthropic answers that the account's 5-hour or weekly window is spent, the proxy remembers that account until its reset time and replays the same request on the next account before Claude Code ever sees the error. To the session nothing happened.

claude ──ANTHROPIC_BASE_URL──▶ proxy 127.0.0.1:48617 ──▶ api.anthropic.com
                                 │  active account's token
                                 └─ 429 "rejected"? mark spent, retry on the next account

Try it without touching any real login: limitbreak video plays a simulated switch in your terminal as a little arcade battle.

Quick start

git clone https://github.com/radu2lupu/limitbreak && cd limitbreak
./install.sh                                   # builds and starts the background service
echo 'source "$HOME/.local/bin/limitbreak-shell.sh"' >> ~/.zshrc   # or ~/.bashrc
limitbreak add                       # sign in another account (browser), repeat per account

Open a new terminal and run claude as usual. When one account's limit is hit, the next one answers.

What it does and does not touch

• Only /v1/messages* requests are sent on an account's token. Everything else (claude.ai connectors, MCP, artifacts, profile) goes through untouched with Claude Code's own credentials, so those keep working.
• It never sets CLAUDE_CODE_OAUTH_TOKEN (that makes Claude Code drop its connectors and local MCP servers).
• Accounts are the logins Claude Code already holds: the default one plus every ~/.claude-<name> config directory that has a login. An account's token is its 1-year token from limitbreak renew when it has one, else the sign-in token read from the macOS Keychain item Claude Code keeps per config dir, or from <dir>/.credentials.json. Tokens are never logged or exposed by GET /_status, and sign-in tokens are never copied.
• An expired token is refreshed by running one tiny claude -p under that account's directory, so Claude Code stays the only thing that refreshes a login.
• If no account can serve a request, it passes through with Claude Code's own login: never worse than no proxy.
• Only a window rejection (429 with anthropic-ratelimit-unified-status: rejected) switches accounts. Ordinary rate limits go back to Claude unchanged.

Install

Needs Go 1.22+ to build, macOS or Linux. Prebuilt binaries are attached to each release.

git clone https://github.com/radu2lupu/limitbreak
cd limitbreak
./install.sh          # builds, installs ~/.local/bin/limitbreak, starts a launchd / systemd user service

Then add the printed line to ~/.zshrc or ~/.bashrc:

source "$HOME/.local/bin/limitbreak-shell.sh"

That defines a claude function that points Claude Code at the proxy when it is running and runs plain claude when it is not. Use ./install.sh --binary <prebuilt> on a host without Go.

Add accounts

limitbreak add            # signs in a new account in your browser, names it account-N
limitbreak add work       # or pick the name
limitbreak list           # accounts, who they are, which one is active (*)
limitbreak remove work    # sign it out and delete its login from this machine

add runs Claude Code's own browser sign-in into its own config directory (~/.claude-<name>), checks the login landed, tells you which email it is, and warns if you added the same account twice. The running proxy picks it up within 30 seconds; no restart. The default account is whatever plain claude auth login signs in. Raw status for scripts: curl -s http://127.0.0.1:48617/_status.

Every switch is logged (~/Library/Logs/limitbreak.log on macOS, the journal on Linux): the account that ran out, when it resets, and the one now serving.

limitbreak list also shows the latest usage headers observed for each account, with their observation time. The proxy stores these snapshots in ~/.local/state/limitbreak/usage.json; they survive restarts. Values are only as fresh as the last response for that account and may be absent if Anthropic did not send them. The headers are undocumented and may change.

Logins that renew themselves

A Claude sign-in ends about 30 days after claude auth login, and refreshing does not move that date. To stop signing in every account once a month, run this once:

limitbreak renew          # every account without a 1-year token (or one ending within 30 days)
limitbreak renew work     # just this one
limitbreak renew --all    # every account, even those with a fresh token

For each account it runs Claude Code's claude setup-token and you authorize once in the browser. The sign-in page opens with the account's email already filled in (as claude auth login --email does). If the browser is already signed in to claude.ai as someone else, that session may win, so check it matches the email renew prints. Each account then gets a 1-year token, kept in the Keychain (item limitbreak-token) on macOS, or in <dir>/.limitbreak-token.json (mode 600) elsewhere and when the Keychain is locked (a Mac reached over SSH). The proxy uses it in place of the sign-in.

To use the same tokens on another machine of yours, sign the same-named accounts in there once, install limitbreak, then from this one run limitbreak share <ssh-host>. Unlike a sign-in, a 1-year token never rotates, so two machines holding it do not log each other out. The other machine checks each token with Anthropic before storing it.

limitbreak list shows when each account's auth ends. The running proxy warns 7 days before any auth ends, in its log and, on macOS, as a notification. If Anthropic rejects an account's token (401), the proxy skips that account for a few minutes and logs run "limitbreak renew <name>"; Claude never sees the error.

Options

limitbreak serve --port 48617 --upstream https://api.anthropic.com --home ~ --credentials auto|keychain|file

A request path may begin with /_c/<client>/ to give a client (an agent, a tool) its own active account; the prefix is stripped before forwarding.

Caveats, read these

• Terms of service. This uses the official claude binary with accounts you own, but spreading load across subscription accounts may conflict with Anthropic's terms for your plan. Read them; you are responsible for your use. Not affiliated with or endorsed by Anthropic.
• Don't copy a login between machines. Refresh tokens rotate; two machines holding the same one will log each other out. Sign in once per machine.
• Headless hosts. On macOS over SSH the Keychain is locked; run the proxy as a user service in the login session (what install.sh does).
• The reset time is read from Anthropic's response headers, which are not a documented contract and may change.

Uninstall

macOS: launchctl bootout gui/$(id -u)/dev.limitbreak; rm ~/Library/LaunchAgents/dev.limitbreak.plist.
Linux: systemctl --user disable --now limitbreak.
Then remove ~/.local/bin/limitbreak* and the source line from your shell profile. Accounts you added stay as ordinary ~/.claude-<name> logins; limitbreak remove <name> deletes one.

Develop

go vet ./... && go test ./...

License

MIT, see LICENSE.