lifeboat

Lifeboat is an open-source tool that provides failover for AI coding agents, licensed under MIT. Its core purpose is: when your AI coding agent encounters rate limits, account bans, or service outages mid-task, it lets the work continue instead of coming to an abrupt halt.

Main capabilities

• Runs tasks on a primary agent (Claude Code by default) and reads the "wall" signals in each response: 429, 529, quota messages, authentication failures, 503.
• When it hits a real wall, it hands the task to the next agent in the failover chain, carrying a context package so the new agent can pick up the work.
• When all cloud agents are unavailable, it falls back to a local model running via Ollama. Local models are slower, private, require no API key, and cannot be banned.
• Context packages are stored in an ordinary folder .lifeboat/ that users own themselves; no matter which provider locks you out, it belongs to you.
• A built-in classifier judges the outcome based on each agent's exit code and output: ok (done, stop), transient (brief fluctuation, retry the same agent a few times), rate_limited/banned/outage (a real wall, hand off to the next agent), unavailable (that agent is not installed, skip). A five-second network blip will not move you off the primary agent; a 429 will.
• Automatic local model selection: Lifeboat reads the machine's memory and, based on the scaling research in docs/RESEARCH.md, chooses a coding model suited to your actual laptop, and will not choose a model the machine cannot handle.

How to use

• It has not yet been published to npm; you need to install it from the repository, and after installation you get the global lifeboat command.
• Point it at any project and run it.
• The doctor command reads the Mac's memory and tells you what you need to install to achieve a true backup.
• You can force a rate limit to trigger and observe the failover process without waiting for a real limit.
• The failover chain drives agents you already trust; each agent is a small adapter, and adding a new adapter requires only one file.
• Use lifeboat config --init to set your own order, then edit .lifeboat/config.json.
• Autonomy and permissions: backup runs are unattended, so they must not stop to ask. By default Claude Code uses acceRunning in ptEdits mode: edit files yourself, do not touch the shell. If the task still needs the shell (such as installing dependencies or running tests), you can grant full autonomy in .lifeboat/config.json, which corresponds to aider --yes-always and cline --yolo; please use it only on trusted projects.

Applicable scenarios

• Building all day on a single agent and a single account, worried that one day being cut off will interrupt work, with no handover and no portability.
• Need resilience across different providers and local hardware, not account evasion.

Limitations and boundaries

• v1 fails over across different providers and to local, but does not do multi-account rotation within the same provider, and will not create new accounts after a ban. The official statement explicitly says that this path violates most providers' terms and is often exactly what causes bans; resilience here means other providers and your own hardware, not account evasion, and this line remains unchanged.
• Difference from model routers: LiteLLM, OpenRouter, and Portkey switch models while the agent stays the same; Claude Code Router and Requesty fail over within a single agent; Lifeboat fails over across the entire agent, brings task context to the new agent, reads bans and quotas rather than just HTTP codes, and ultimately falls back to a local model that cannot be cut off.

Current availability status

• Claude Code as the primary agent can run real tasks headlessly and edit files, verified end to end.
• The failover engine (detecting walls, handover, carrying context packages, retrying fluctuations, clean exhaustion) has test coverage.
• Local fallback communicates through Ollama's OpenAI-compatible API, with test coverage based on a stub server.
• The Aider and OpenCode adapters run their documented headless commands; install either one and set the key to join the chain. If the command or model is wrong, Lifeboat will report the error and continue, and will never fake success.
• To observe a real cloud-to-local handover, install Ollama, pull a model, and then force a limit to be triggered.

Roadmap (officially listed)

• Live view showing the handover happening.
• Menu bar indicator light: green on Claude, yellow for local backup.
• Context packages summarize not only the task, but also the diff so far.
• More adapters (Codex, Cline, Gemini) asplug-and-play files.