Claude Failover — AI skill for Claude Code

claude-failover

**A backup brain for Claude Code. When your Claude Max plan hits its limit, your tokens are exhausted, or Claude is down — flip one command and your `claude -p` agents keep running on a local LLM. Default to Claude. Local only when you need it.**

Built for people who **like** Claude and want to keep using it as primary, but don't want a usage cap or an Anthropic outage to halt every script that depends on `claude -p`.

This is **NOT** "ditch Claude and run everything local." For that, see [claude-code-local](https://github.com/nicedreamzapp/claude-code-local) — that project replaces Claude entirely with a local model. This project is the opposite angle: **Claude stays primary**, local is just there for the day you need it.

---

When you'd want this

• You're on Claude Max ($100/mo SDK credit, or any tier with a usage cap) and you sometimes hit it mid-month
• You run headless claude -p scripts (cron jobs, agents, watchers, custom tooling) and don't want them to die during an Anthropic outage
• You want to run lower-stakes work locally to preserve your Claude budget for the work that actually needs Claude's quality
• You have an Apple Silicon Mac with enough RAM to run an MLX model (8 GB minimum for a 7B-class model, 32+ GB for 30B class)

How it works

┌─────────────────────┐
│ Your script         │
│ subprocess.run([    │       reads ~/.local/state/llm-backend
│   "agent-llm",      │ ─────────────────────────────────────────┐
│   "-p", prompt      │                                          │
│ ])                  │                                          │
└─────────────────────┘                                          │
                                                                 ▼
                                          ┌─────────────────────────────────────┐
                                          │ flag = "claude"  →  exec real `claude -p`│
                                          │ flag = "local"   →  POST to loca