Tendrl

AI Provider

An AI Provider connects Strand to an AI vendor. It holds the vendor's API key and a few defaults. AI Agent nodes use it to send prompts to that vendor's models.

Direction: Read / Write | Types: ai.openai, ai.anthropic, ai.google, ai.deepseek, ai.openrouter, ai.openai_compatible

Vendors#

Vendor Get an API key Default model
OpenAI platform.openai.com/api-keys gpt-4.1-mini
Anthropic Claude console.anthropic.com/settings/keys claude-sonnet-5
Google Gemini aistudio.google.com/apikey gemini-3.5-flash
DeepSeek platform.deepseek.com/api_keys deepseek-chat
OpenRouter openrouter.ai/keys none: you choose the model
OpenAI-compatible From the service you're connecting to none: you choose the model

OpenRouter gives one API key for models from many vendors. Enter the model as OpenRouter names it, such as anthropic/claude-sonnet-4.5 or openai/gpt-4.1-mini (see openrouter.ai/models). Strand still shows only the settings that model accepts.

OpenAI-compatible covers any service that speaks the OpenAI Chat Completions API: a model you host behind vLLM, an LLM gateway, or an internal deployment. Enter its API root as the Base URL (for example https://llm.example.com/v1). Strand is hosted, so the endpoint must be reachable from the internet. Private, local and internal addresses are refused when you save.

You can add several providers, for example one per vendor, or separate keys for production and testing.

Setup#

  1. On the Connectors page, choose the vendor under AI Providers & MCP.
  2. Enter a name and the API key (and, for OpenAI-compatible, the Base URL and a model).
  3. Click Test to check the key with the vendor.

Configuration fields#

Field Required Default Description
Name Yes - Friendly name, shown when you pick a provider on an AI Agent node
API Key Yes - The vendor API key, encrypted at rest
Base URL OpenAI-compatible only - The endpoint's OpenAI-compatible API root. OpenAI providers can also set one to route through a gateway
Default Model No (OpenRouter and OpenAI-compatible: yes) vendor default Model AI Agent nodes use unless they choose another
Temperature No 0.7 Response randomness. Ignored by models that don't accept it
Max Tokens No 4096 Maximum output tokens, including any reasoning before the answer
Organization ID No (OpenAI only) - OpenAI organization for billing attribution
Timeout No 120 Request timeout in seconds (10–600)

Models#

On an AI Agent node, the model list shows the models your provider's key can use, asked from the vendor and refreshed hourly. Models the vendor has retired, or doesn't offer your account, aren't listed. If the vendor can't be reached, the list falls back to every model Strand knows for that vendor. Vendors sometimes list a model they then refuse for your key; when a run gets that refusal, Strand hides the model from that provider's list for 7 days, or until you edit the provider.

Only current names are shown: a dated version such as gpt-4o-2024-08-06 is hidden when gpt-4o is listed. The vendor prefix is dropped from labels (sonnet-5-5 rather than claude-sonnet-5-5), but Strand stores the full model ID. For OpenRouter and OpenAI-compatible providers, the list is whatever the service offers.

To use a model that isn't listed, such as one released after the list was updated, choose Custom model... and enter its full ID. A node saved with a model that's no longer listed keeps it, shown as a custom model.

Models that can't work in an AI Agent node, such as audio, image-generation and embedding models, aren't listed.

What a model supports#

You don't configure per-model behavior. When an AI Agent node picks a model, Strand shows only the settings that model accepts:

This comes from the model data in pydantic-ai, the library Strand uses to talk to every vendor, so it follows new models as that data is updated.

← Back to Strand docs