Nicholas Saunders

limit-handoff

A free, open-source Claude Code plugin. When a chat stops at a usage limit, carry on in a new chat without explaining the task again.

Builder and maintainerSince Nov 2026github.com ↗

The problem

I build with Claude Code and use more than one Claude account. When a long task hit a usage limit, the chat stopped. The limit belongs to the account, so a new chat on the same account was blocked too.

A chat on my other account knew nothing. I spent time, and the new account’s tokens, explaining the goal, what was done, which files had changed and where I had stopped. Something always got missed.

What I built

limit-handoff is a plugin for Claude Code. The moment a chat stops at a limit, it:

  • saves your place in a HANDOFF.md file in your project folder: the goal, what is done, the next steps, and a note about the cut-off;
  • copies a short “paste prompt” to your clipboard, and also saves it as PASTE-PROMPT.txt in the project folder;
  • shows a desktop notification.

On your other account, you open a new chat in the same project folder and paste. Claude reads the notes, checks for a half-finished step, and carries on from the next one.

It also catches chats that ended before it could run, and reminds the new chat about an open handoff. Everything stays on your computer. It makes no network calls, and it does not change or get around any limit. It only saves notes.

It has 501 automated tests on macOS and Linux, and it is free under the MIT licence.

Try it

You need Claude Code 2.1.78 or newer and jq, on macOS or Linux. In Terminal:

claude plugin install limit-handoff --marketplace nicholas-a-saunders/limit-handoff

Then restart Claude Code. That is all the setup.

It is an independent project, not made by Anthropic. Use each of your accounts within Anthropic’s terms.

What I learned

  • Test it the way a stranger would. A full first-time-user test, run before launch, found things I had missed: a command that only works inside a Claude chat, not in Terminal, and lines that promised more than the tool does.
  • Say what it does not do. The clearest line on the page is the one that says the tool does not get around any limit.
  • I built it with Claude Code, and used AI agents to review and test it. The agents found real bugs. Every fix still needed checking by hand.

What’s next

Feedback from people using it. Windows support is not tested yet. Bugs and ideas are welcome as issues on GitHub.