Models and providers

AgentRow calls models through provider configurations set up per organisation under Settings → Provider settings. Supported providers: OpenRouter, OpenAI, Anthropic, vLLM (self-hosted OpenAI-compatible servers) and Ollama. Each provider configuration carries its own API key or endpoint.

Row is the model AgentRow provides with the paid plans (Plans, pricing and what's next): switch it on and the assistant needs no provider keys of its own. Until then, and whenever you prefer, use your own provider configurations.

#Provider keys and base URLs

• OpenRouter, OpenAI and Anthropic need an API key; a Base URL is optional.
• vLLM and Ollama need no key but must have a Base URL: the address of your own model server. Saving one without it is refused ("base_url is required for vllm and ollama providers"), and a saved base URL can be replaced but not cleared.
• AgentRow has no built-in endpoint for vLLM or Ollama. A model on one of them with no base URL fails its run with PROVIDER_BASE_URL_REQUIRED before any model is called.
• A base URL must be a public http or https address. localhost and private networks are refused when you save and again on every request (Guardrails, approvals and safety).
• Keys are stored encrypted and never shown again.

#Model settings

Per agent, and per LLM node in a workflow, you choose:

• Provider and model from the organisation catalogue. A workflow node can use a different model than the agent default.
• Temperature (0 to 2) and max output tokens.
• Context limits: model context limit, max input tokens and a safety margin. The runtime trims history and knowledge to fit the budget and reports what it trimmed.
• Streaming: answers stream token by token to the chat and the widget.
• Response format: free text or JSON.
• Tool choice: let the model decide, force a tool, or disable tools for that call.
• Cost limit per run, fallback models if the primary fails, and stop sequences.
• Prompt versions and labels so a node can pin a specific version of a prompt.

#The Model tab and the LLM node

The agent's Model tab sets its default model. An LLM node in the workflow uses that model unless a provider and model are set on the node itself; a model set on the node overrides the Model tab for that step only.

A new agent's LLM node carries no model of its own, so changing the Model tab changes the model the agent runs. In the node's settings on the Workflow tab, the provider and model show as inherited from the agent (Agent defaults). Override sets a value on the node; Inherit returns it to the agent's setting, and inheriting the provider releases the node's model with it.

An agent created earlier may still have a model stored on its LLM node, which then wins over the Model tab. If changing the Model tab has no effect, create a draft of the workflow, open the LLM node, choose Inherit on the provider (or set the model you want there) and publish.

#Embedding providers

Embeddings for knowledge bases, memory and tool selection use embedding provider configurations, managed under Knowledge → Embedding providers:

• AgentRow Embedding is hosted by AgentRow and needs no endpoint, model or timeout settings.
• Self-hosted is an Ollama server at your own base URL, with its model, dimensions, batch size and timeout. Its base URL must be a public address: a localhost or private address is refused (Guardrails, approvals and safety).

You pick the embedding provider when you build a knowledge base index, or in the node settings.