Radu Lupu has released Limitbreak, a local proxy that allows a single Claude Code installation to rotate between multiple Anthropic accounts automatically. The tool intercepts requests to api.anthropic.com and, upon receiving a 429 rate-limit rejection indicating a spent window, marks the current account as exhausted and replays the identical request using the next available account. This happens transparently, meaning the active coding session never sees an error and does not need to be restarted.

Under the Hood

The architecture is deliberately minimal: a single Go binary using only the standard library consumes just a few megabytes of RAM. It listens on port 48617 and routes only /v1/messages* endpoints through the rotating tokens. Crucially, the proxy avoids setting CLAUDE_CODE_OAUTH_TOKEN, which would otherwise strip Claude Code of its connectors and local MCP server support. Instead, it reads tokens directly from the macOS Keychain or local credential files associated with each ~/.claude- directory, ensuring that non-message traffic continues to use the primary login's credentials.

Operational Mechanics

Installation involves cloning the repository and running install.sh, which builds the binary and starts a background service via launchd on macOS or systemd on Linux. Users add accounts through a browser-based sign-in flow that creates isolated config directories. The proxy polls for new accounts every 30 seconds, so no service restart is required when adding capacity. If all configured accounts are exhausted, the system fails open, passing the request through with the default login to ensure the developer is never worse off than they would be without the proxy.

Terms of Service and Caveats

The README explicitly warns that spreading load across multiple subscription accounts may violate Anthropic's Terms of Service for specific plans. While the tool uses the official binary, the behavior of rotating sessions to evade usage caps exists in a gray area. Additionally, the reset time logic relies on Anthropic's undocumented response headers, which could change without notice. Users are advised to read their plan's terms carefully, as Limitbreak is not affiliated with or endorsed by Anthropic.

Key Takeaways

  • Limitbreak uses a lightweight Go proxy to swap Claude accounts mid-session when 5-hour or weekly limits are hit.
  • The tool preserves MCP and connector functionality by avoiding the CLAUDE_CODE_OAUTH_TOKEN environment variable.
  • Reset times are derived from undocumented API headers, introducing potential fragility if Anthropic changes their response schema.
  • Users must verify their specific subscription terms, as rotating accounts to bypass limits may violate Anthropic's ToS.

The Bottom Line

Limitbreak is a clever engineering hack that solves a real pain point for power users, but its reliance on undocumented API headers and potential ToS violations makes it a tool for enthusiasts rather than production-critical workflows.

Sources

https://github.com/radu2lupu/limitbreak

> From The Wire