Ollama and local models
Run tasks on local models with Ollama or LM Studio — no key, nothing leaves the machine — and combine them with a network-off sandbox.
Ollama
- Install Ollama on the worker machine and pull a model that supports tool calls, for example
ollama pull qwen2.5-coder:14b. - Worker UI → AI models → Add → Ollama (local). Default address
http://127.0.0.1:11434. - Optionally list the models; otherwise every model Ollama reports (
/api/tags) is used.
Codex, OpenCode and Aider use Ollama directly; Claude Code and Gemini CLI through the gateway (Ollama's /v1).
LM Studio
Start LM Studio's local server, then add LM Studio (local) (default http://127.0.0.1:1234/v1) with the model id
LM Studio shows. All agents reach it through the gateway.
Fully local tasks
On Linux and macOS workers, the sandbox policy can cut network access entirely:
Project policy
{ "sandbox": { "mode": "required", "network": false } }The agent can then reach only local services such as the model. Use this only with local models — a hosted model API would be unreachable.