What is claude-failover?

claude-failover is a backup routing layer for the Claude Code CLI. When you hit your monthly usage cap or Anthropic's service is temporarily down, you flip a single command and all your scripts that call Claude automatically switch to a local AI model running on your own machine, no code changes required. When Claude is available again, one more command flips everything back. The way it works is through a lightweight shim called agent-llm that you drop into your PATH. Instead of calling the Claude CLI directly in your scripts, you call agent-llm. Every time it runs, it reads a flag file on your computer to decide where to send the request: if the flag says "claude," it passes the call straight to the real Claude CLI as if nothing changed, if it says "local," it sends the request to a locally running AI server instead. You switch modes with llm-failover local or llm-failover claude. The local model is lazy-loaded, meaning it only occupies memory when you are actually in failover mode. Flipping to local mode starts the server and loads the model (which can take 30 to 90 seconds the first time depending on your machine). Flipping back unloads it and frees the RAM. If you want instant failover without the warm-up delay, you can configure the server to stay loaded continuously. You would use this if you run automated scripts or cron jobs that depend on the Claude CLI and you cannot afford those jobs to stop when you hit a usage limit or during an outage. It requires an Apple Silicon Mac, Python 3.9 or later, the Claude Code CLI, and the mlx-lm package. The full README is longer than what was provided.