What it does
Lifeboat runs a coding task on a primary agent, Claude Code by default, and reads each response for signs that the agent has been cut off: rate limits, quota messages, authentication failures, bans or outages. When that happens it passes the same task to the next agent in your chain along with a context bundle, so the new agent continues the work instead of starting from nothing. The chain can include Aider and OpenCode and ends at a local model served through Ollama, which the README describes as needing no API key and being impossible to ban. A doctor command checks which agents are installed and suggests a local coding model sized to your Mac's memory. Short blips are retried on the same agent rather than triggering a handoff. The bundle lives in a plain .lifeboat folder in your project. The README is explicit about limits: Lifeboat fails over between different providers and down to local hardware, and it does not rotate accounts on one provider or create new accounts after a ban. It is not on npm yet and is installed by cloning the repository. Claude Code as the primary agent is reported as verified end to end, while the Aider and OpenCode adapters run their documented headless commands.
Capabilities
- Failover chain across Claude Code, Aider, OpenCode and a local Ollama model
- Sorts each result as ok, transient, rate limited, banned, outage or unavailable
- Carries a context bundle in a plain .lifeboat folder to the next agent
- doctor command that checks installed agents and suggests a local model for your RAM
- Simulated rate limit for testing the handoff without waiting for a real one
When to use it
- Long coding sessions on one agent subscription that could hit a usage cap
- Wanting a private, offline fallback for when every cloud agent is unavailable
- Testing how your setup behaves when a provider fails
