AIs of an organization and the projects that use them

The assistant, the project draft of the new project wizard and Edit with AI run on the AIs an organization defines. Each AI is a provider, a model and the credentials to reach it. An organization may define several, for example Claude for every team and GPT on the Azure subscription of one department, and the user picks one per request.

Add an AI

In the organization settings (/organization), the AI tab lists under Artificial intelligences the AIs of the organization: every member sees the list; owners and admins add, edit and delete AIs with Add AI.

Provider in the form	Model	Configuration	Model field
Claude (Anthropic)	Claude through the Anthropic API	API key	Optional, claude-opus-5-5 when empty
Claude (Azure AI Foundry)	Claude deployed on Microsoft Foundry	Resource, API key	The deployment, optional, claude-opus-5-5 when empty
Claude (AWS Bedrock)	Claude through Amazon Bedrock	Region, access key ID, secret access key	Required: the Bedrock model id or inference profile
GPT (Azure OpenAI)	GPT deployed on Azure OpenAI	Endpoint, API key, Reasoning model	Required: the deployment
Gemini (Google)	Gemini through the Gemini API	API key	Required, for example a current Gemini Pro

Every AI also has:

• a Name, unique in the organization whatever the case;
• an optional Maximum effort, the highest reasoning effort users can pick with it;
• Default AI: preselected in every project that can use it. At most one AI of the organization is the default, so marking one unmarks the others.

Credentials

API keys and secret access keys are stored encrypted, like the storage credentials (see Encryption at rest), and never shown again. When you edit an AI, leave a hidden value as it is to keep it. A stored key is kept only for the provider, resource, endpoint and access key it was saved with: point the AI somewhere else and the form asks you to type the key again, so a key is never sent where it was not meant to go.

Deleting an AI removes it from every project restricted to it. Deleting the organization deletes its AIs.

Provider notes

Claude (Anthropic)

Requests that Claude declines are retried on another model by the Anthropic API itself.

Claude (Azure AI Foundry)

The Resource is the name of the Foundry resource, as in https://<resource>.services.ai.azure.com, and the model is the name of the Claude deployment. Foundry has no server-side fallback: a declined request is reported as declined.

Claude (AWS Bedrock)

Requests go to the Messages API of Bedrock in the region you set, signed with the access key of the form, never with the AWS credentials of the server. The model is a Bedrock model id such as anthropic.claude-opus-5-5, or an inference profile such as global.anthropic.claude-sonnet-5-5; the key needs permission to invoke it. Effort, reasoning, prompt caching and structured outputs work as on the Anthropic API.

GPT (Azure OpenAI)

The Endpoint is https://<resource>.openai.azure.com/ and the model is the name of the GPT deployment. Requests go to the v1 endpoint through the Responses API, so there is no api-version to set: Chat Completions rejects function tools with reasoning on from GPT-5.6 onwards, which is why the Responses API is used. Nothing is stored on Azure: the conversation is sent whole on every turn.

Turn Reasoning model on for reasoning deployments (GPT-5 and later, o-series): the effort is then sent to the model. Leave it off for other models, which reject it.

The endpoint must be an Azure host: https://<resource>.openai.azure.com/, .cognitiveservices.azure.com or .services.ai.azure.com, with no IP address, no credentials, no port other than 443 and no path. The server sends the key there, so an arbitrary URL would let an organization administrator reach the network the server runs in. For the sovereign clouds a self-hosted installation adds suffixes with AZURE_OPENAI_ENDPOINT_SUFFIXES, comma separated, for example .openai.azure.us,.openai.azure.cn.

Gemini (Google)

The model has no default, because Gemini models are retired often. Use a model with function calling, such as a current Gemini Pro. Gemini ignores the reasoning effort.

:::note Errors from the provider A provider's own error message is never shown in the console: you get a generic error, or a rate limit notice, and the provider's message goes to the server log. Edit with AI retries by itself when the model hits its rate limit. :::

Which AIs a project may use

By default a project may use every AI of its organization, including those added later. Under Project access in the same tab, owners and admins choose, project by project:

• All the AIs of the organization (the default), including those added later;
• Only the selected AIs: an AI added later is not included;
• no AI at all: the assistant and the AI features are turned off in that project, and nothing is sent to any provider.

The AI preselected in a project is the organization's default when the project may use it, otherwise the first one available. A request naming an AI the project may not use is refused before anything starts. A follow-up in Edit with AI runs on the AI of the previous message unless you pick another one: switching AI in the middle of a conversation is safe, because the conversation holds nothing specific to a provider.

The new project wizard runs before the project exists, so it uses the AIs of the organization the project is created in.

The system AI

A self-hosted installation can describe one more AI in its environment variables: the System AI, offered to every organization next to its own AIs. It is read-only in the console and can be left out of a project like any other. Without these variables there is no system AI.

AI_PROVIDER	Model	Required variables	Optional
anthropic	Claude, Anthropic API	ANTHROPIC_API_KEY	ANTHROPIC_MODEL (claude-opus-5-5)
azure-foundry	Claude on Microsoft Foundry	AZURE_FOUNDRY_RESOURCE, AZURE_FOUNDRY_API_KEY	AZURE_FOUNDRY_MODEL (claude-opus-5-5)
aws-bedrock	Claude on Amazon Bedrock	AWS_BEDROCK_REGION, AWS_BEDROCK_ACCESS_KEY_ID, AWS_BEDROCK_SECRET_ACCESS_KEY, AWS_BEDROCK_MODEL
azure-openai	GPT on Azure OpenAI	AZURE_OPENAI_ENDPOINT, AZURE_OPENAI_API_KEY, AZURE_OPENAI_DEPLOYMENT	AZURE_OPENAI_REASONING (false)
gemini	Gemini, Gemini API	GEMINI_API_KEY, GEMINI_MODEL

• AI_PROVIDER left empty with ANTHROPIC_API_KEY set means Claude through the Anthropic API.
• A provider named without one of its required variables stops the backend at boot, with the missing variables in the error.
• AI_MAX_EFFORT (low, medium, high, xhigh or max) is the maximum effort of the system AI.
• ASSISTANT_RATE_LIMIT_MAX (default 20) caps the questions to the assistant per minute; Edit with AI requests count against it too.

# docker-compose: Claude through the Anthropic API as the system AI
environment:
  AI_PROVIDER: anthropic
  ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY}
  AI_MAX_EFFORT: high

Authentication is by API key on every provider (access key on Bedrock).