Skip to main content
A run needs three kinds of credential: a harness to drive the agent, a sandbox to run it in, and, for a case that states a human, a model to play them. Each one is a connector.

​
Adding one

add takes the integration and prompts for the secret:
There is no --key option, and that is deliberate. A secret in a flag is kept in shell history and readable from the process list by anyone on the machine for as long as the command runs. It is prompted for instead. In a script, pipe it:
The value is encrypted when it arrives and is never returned. list states that a connection exists, not what is in it.

​
Model providers

openai, anthropic, google, xai, moonshotai, deepseek, groq and openrouter all speak the same chat-completions shape, so a case’s human runs on whichever one you connect. The first connected provider is the one it uses. A model judge calls OpenAI’s responses API, which the others do not serve, so a case scored by one needs openai specifically.

​
Choosing the auth method

An integration that accepts more than one way in takes --auth:
Run anpord connectors integrations to see the methods each one accepts and the fields they need. A method that signs in through a browser cannot be completed from the terminal, and the command says so rather than failing.

​
Which connector a run uses

--default marks the one runs reach for:

​
Removing one

A run already holding the credential finishes; the next one cannot start without another connector for that integration.