Keep tasks running when an agent hits its limit
Add add-on models on a worker so a task continues from its checkpoint on another provider when the agent's own login reaches its usage limit.
By default every task runs on the agent's own login — your Claude subscription, your ChatGPT login for Codex, and so on. When that login reaches its usage limit, the task waits for the reset. Add-on models let it continue instead.
Add a provider on the worker
Worker UI → AI models → Add. Good first choices:
- OpenRouter — one key for many models; you can Sign in with OpenRouter instead of pasting a key.
- Ollama or LM Studio — local models, no key, nothing leaves the machine.
- Groq, DeepSeek, NVIDIA NIM, OpenAI, Google Gemini, Anthropic — with an API key.
List the models to use, best first (for example qwen/qwen3-coder:free). Models must support tool calls to
drive a coding agent. Keys go to the operating system's credential store.
Choose what happens at the limit
Worker UI → AI models → When a harness reaches its limit:
- Ask me (default): the task waits in the dashboard; reply
switchto continue on add-on models orwait. - Switch automatically: the task continues at once on the first available add-on model.
Order the providers
Add-on providers are tried in list order; move them up and down. Inside one turn, a request that hits a limit, an outage or a refusal moves to the next model, as long as nothing has been streamed yet.
How it works
The worker's model gateway sits between the agent and the add-on provider. It listens on 127.0.0.1
only, gives each agent session its own token, and translates each agent's API — Anthropic Messages for Claude
Code, the Responses API for Codex, the Gemini API for Gemini CLI, chat completions for OpenCode and Aider — to
the provider's OpenAI-compatible API and back, including tool calls. See Model gateway.
The next session of the task goes back to the agent's own login once its limit has reset.
What you lose on add-on models
Thinking blocks, prompt caching and hosted tools such as web search are not available through the gateway.