Models

The assistant block picks the model. Leave the block out and there’s no assistant.

Two providers are built in: openai-compatible, the default, and anthropic. The first covers anything that speaks the OpenAI chat API: Ollama, llama.cpp, vLLM, LM Studio, OpenRouter, OpenAI.

The model needs to support tool calling. Without it, the assistant can’t read anything.

Local, through Ollama

config.yaml

assistant:
  name: Valet
  baseURL: http://ollama:11434/v1
  model: qwen3:8b

If you leave out baseURL, it defaults to http://localhost:11434/v1. Inside the container, localhost is the container itself. So set baseURL to wherever Ollama really is, such as the Compose service name.

Nothing leaves your network with this setup. The model runs on your hardware, and Valet talks to it directly.

Other OpenAI-compatible servers

config.yaml

assistant:
  baseURL: http://llama:8080/v1
  apiKey: ${LLAMA_KEY}        # only if the server wants one
  model: local-model

A hosted one looks the same, with its own baseURL, apiKey and model.

Anthropic

config.yaml

assistant:
  provider: anthropic
  apiKey: ${ANTHROPIC_API_KEY}
  model: your-model-name

baseURL is optional here.

What a hosted model sees

A hosted model sees what the tools return: card data, document text if you let it read documents, and logs if you let it read logs. Keys don’t go to it. Valet redacts them first, but it can only redact values that came from a ${NAME} in your config. Everything else in a log or a document goes as it is.

If that matters to you, use a local model, or set sensitive integrations to none.

Options

Option	Default	What it does
provider	openai-compatible	openai-compatible or anthropic.
baseURL	http://localhost:11434/v1 for openai-compatible	Where the server is.
apiKey	none	Use ${NAME}.
model	required	The model’s name on that server.
name	Assistant	What it’s called on the page.
instructions	none	Extra text for the system prompt.
access	{ default: read }	What it may use. See access levels.
maxSteps	6	Tool rounds per answer, 1 to 20. The last step is kept for answering.
background	see below	What it does on its own.
suggestions	four questions	Offered on an empty chat and in the palette.

The chat window has its own options under chat. Only the last 24 messages go to the model, and tool results from older turns shrink to a preview, so a long chat doesn’t get slower with every turn.

Background use

Some features call the model without being asked. They get no tools, so they can only say something.

Option	Default	What it does
background.watch	true	When two or more things are wrong at once, asks whether one cause explains them.
background.brief	true	A short summary, shown when you hover the status sentence and rewritten at most every half hour.
background.suggest	true	Questions worth asking right now.
background.perDay	100	The most background calls a day, from 1 to 500.

Calls go one at a time, so a single GPU isn’t asked to do two things at once. Rules decide when a call is worth making: one problem on its own never reaches the model. To stop them all, set the first three to false.

Handing the chat to another agent

assistant.agent sends the chat to another agent with an OpenAI-compatible API, such as Hermes.

config.yaml

assistant:
  model: qwen3:8b
  agent:
    name: Hermes
    baseURL: http://hermes:8000/v1
    apiKey: ${HERMES_KEY}
    model: hermes

Access rules don’t apply

The agent gets the conversation and nothing else from Valet: no tools, no keys, no data. But it runs its own tools, and Valet can’t see or stop them. Access levels and approvals protect Valet’s assistant, not this one.

The chat says so at the top. Answers are plain text, with no cards. A read-only page stays on Valet’s own assistant, because the agent can act and that page may only look.