Choosing an LLM provider

By default, the backend uses OpenAI (OPENAI_API_KEY in .env). If you’d rather use a different provider — your own Azure OpenAI deployment, AWS Bedrock, Anthropic directly, or something else — it’s configurable.

You don’t have to redeploy to switch providers. This is also configurable live from Django Admin (Agent Studio → Coding agent runtime configs), which takes precedence over .env once you enable a config there. Use .env for the version you want to ship with by default, and Django Admin if you want to change it afterward without restarting containers.

Two tracks

RELAI_AGENT_PROVIDER picks which track you’re on:

codex (the default) — an OpenAI-compatible track. Works with OpenAI, Azure OpenAI, AWS Bedrock, or any OpenAI-compatible gateway.
claude — an Anthropic-compatible track. Works with direct Anthropic, or Claude on AWS Bedrock, Google Vertex AI, AWS Marketplace, or Microsoft Foundry.

Within whichever track you pick, RELAI_AGENT_MODEL_PROVIDER selects the concrete backend. The sections below cover each one.

Nothing extra to configure beyond what’s already in Configuration:

OPENAI_API_KEY="sk-..."

RELAI_AGENT_PROVIDER="codex"
RELAI_AGENT_MODEL_PROVIDER="azure"
RELAI_AGENT_MODEL="<your-azure-deployment-name>"
RELAI_AGENT_BASE_URL="https://<your-resource>.openai.azure.com"
RELAI_AGENT_API_KEY_ENV="AZURE_OPENAI_API_KEY"
RELAI_AGENT_QUERY_PARAMS={"api-version": "<azure-api-version>"}
AZURE_OPENAI_API_KEY="..."

RELAI_AGENT_MODEL_PROVIDER doesn’t have to be literally azure — any custom value works, it’s just a label. What matters is RELAI_AGENT_BASE_URL pointing at your deployment endpoint and the API version going through RELAI_AGENT_QUERY_PARAMS.

Native Codex on Bedrock:

RELAI_AGENT_PROVIDER="codex"
RELAI_AGENT_MODEL_PROVIDER="amazon-bedrock"
RELAI_AGENT_MODEL="openai.gpt-5.4"
RELAI_AGENT_AWS_REGION="<your-aws-region>"

Leave RELAI_AGENT_API_KEY_ENV unset. For SDK-based auth, use an AWS profile or the default AWS credential chain; for Bedrock API-key auth, set AWS_BEARER_TOKEN_BEDROCK in the container environment instead.For an AWS-hosted OpenAI-compatible gateway instead of native Bedrock:

RELAI_AGENT_PROVIDER="codex"
RELAI_AGENT_MODEL_PROVIDER="aws"
RELAI_AGENT_BASE_URL="https://<your-gateway-endpoint>"
RELAI_AGENT_API_KEY_ENV="<YOUR_GATEWAY_KEY_ENV>"

RELAI_AGENT_PROVIDER="claude"
RELAI_AGENT_MODEL="claude-sonnet-4-6"
RELAI_CLAUDE_API_KEY_ENV="ANTHROPIC_API_KEY"
ANTHROPIC_API_KEY="sk-ant-..."

Claude on AWS Bedrock:

RELAI_AGENT_PROVIDER="claude"
RELAI_AGENT_MODEL_PROVIDER="bedrock"
RELAI_AGENT_MODEL="<bedrock-model-id>"
RELAI_AGENT_AWS_REGION="<your-aws-region>"

Claude on Google Vertex AI:

RELAI_AGENT_PROVIDER="claude"
RELAI_AGENT_MODEL_PROVIDER="vertex"
RELAI_AGENT_VERTEX_PROJECT_ID="<your-gcp-project>"
RELAI_AGENT_VERTEX_REGION="<your-gcp-region>"

Claude on Microsoft Foundry:

RELAI_AGENT_PROVIDER="claude"
RELAI_AGENT_MODEL_PROVIDER="foundry"
RELAI_AGENT_FOUNDRY_RESOURCE="<your-foundry-resource>"
# Leave RELAI_CLAUDE_API_KEY_ENV unset to use Entra ID/gateway auth,
# or set it to an env var containing your Foundry API key.

Codex-only, via a local model runner:

RELAI_AGENT_PROVIDER="codex"
RELAI_AGENT_OSS="true"
RELAI_AGENT_LOCAL_PROVIDER="ollama"  # or "lmstudio"

Not every combination is valid — for example, the claude track only accepts anthropic, bedrock, vertex, anthropic-aws, or foundry as a model provider, and the codex track rejects those same values. If a config is invalid, the backend will tell you exactly what’s wrong when it starts up or when you try to enable it in Django Admin.

​Two tracks

Two tracks