Engines
Valhalla ships no model and runs none of ours. It drives one of four engines, and choosing one is the only real decision in setting it up.
The four engines
| Engine | Needs | Gives you |
|---|---|---|
| claude-cli (default) | Claude Code installed and signed in with a Claude subscription. Run claude once in a terminal; Valhalla finds it where it was installed. | Everything: members can read and edit files, run commands and show their reasoning. Your subscription pays for it. |
| anthropic-api | An API key in the ANTHROPIC_API_KEY environment variable. | Answers, with reasoning. No tools. |
| openai | An OPENAI_API_KEY, or an OpenAI-compatible local server such as Ollama (set the base URL in Settings › Engine; a local address needs no key). | Answers from any compatible model. |
| echo | Nothing. | No model at all: it returns the assembled request, which is how to see exactly what would be sent. |
Choose one in Settings › Engine, which also shows which engines are ready.
Model routing: auto
The model is auto by default. Each turn is routed to a fast, balanced or deep model from what the request asks for and carries, and the reason is one sentence you can read. Settings › Engine lets you pin one model instead, and a single conversation can override it from the model picker beside the composer.
Effort and reasoning
Effort (Settings › Engine) sets how hard the model thinks. With no effort set the claude-cli engine shows no reasoning at all; high is a good default.
A local model
Choose the openai engine and set the base URL to your server, for example Ollama's http://localhost:11434/v1. A local address needs no key. Members answer with whatever the model can do; tools are a claude-cli feature.
What the engine sends
Your message and what Valhalla assembled around it (the members' voices, the matched skills and knowledge, and in a repository its branch and diff) go to the engine's provider under that provider's terms. Nothing goes to us. The echo engine shows you that request without sending it anywhere.