Environment Variables¶

Dev Agents is configured through config.yaml, and almost every value there is a Jinja template that reads an environment variable. In practice you only edit a .env file.

The .env file is loaded from the current working directory at startup (.env.example in the repository is the annotated template). Only variables that a template actually references have an effect — see config.yaml for how to add your own.

Minimum setup¶

# One LLM provider
ANTHROPIC_API_KEY=sk-ant-...

# Models (Pydantic AI strings: provider:model-name)
LLM_MODEL_LARGE=anthropic:claude-sonnet-5
LLM_MODEL_SMALL=anthropic:claude-haiku-4-5

# Local checkout to analyze
GIT_REPO_PATH=/code

LLM providers¶

Credentials are read by the Pydantic AI provider itself, not by Dev Agents.

Any provider supported by Pydantic AI works; set that provider's own credential variables.

Core¶

Verbose console logging is enabled with the -v flag (which sets DEV_AGENTS_CONSOLE_LOGGING=1 for all services).

Git repository¶

See Git integration.

Providers¶

Every provider has a *_MOCK flag that returns canned data instead of calling the API.

Clone URLs used with a provider token must be https:// and point at the host of the configured provider URL (AZURE_URL, GITLAB_API_URL, or the web host derived from GITHUB_API_URL / BITBUCKET_API_URL). The *_ALLOW_INSECURE_CLONE_URL flags (default False) are development-only and additionally permit http://; the host check still applies. See config.yaml.

Linear has no environment templates in the bundled defaults — configure it in config/config.custom.yaml, see Linear.

Slack¶

Both tokens must be set for the Slack entrypoint to start. See Slack.

Platform services¶

Email, Mattermost and Telegram each start only when <NAME>_ENABLED is truthy and their credentials are present. Without an allowlist, every user on that platform can drive the agent. See platform services.

HTTP entrypoints¶

AG-UI and the OpenAI-compatible API share one server, which starts if either is enabled.

OPENAI_* here configures the inbound API. The OpenAI models you call are set with LLM_MODEL_* and OPENAI_API_KEY.

Both entrypoints run the agent, so authentication is fail-closed. An enabled entrypoint whose AGUI_API_KEYS / OPENAI_API_KEYS is empty refuses to start with a configuration error naming the variables to set. Running without authentication is an explicit opt-in via AGUI_ALLOW_UNAUTHENTICATED=true / OPENAI_ALLOW_UNAUTHENTICATED=true, which is only appropriate when the port is reachable from a trusted network alone; the server logs a warning at startup when it is set. Clients authenticate with Authorization: Bearer <key>.

CLI, NATS and subagents¶

Security¶

• Keep .env out of version control and restrict its permissions (chmod 600 .env).
• Use read-only tokens where the provider supports them.
• Prefer your platform's secret store (Docker secrets, Kubernetes secrets) over a plain .env in production.

Next steps¶

• config.yaml — the structure behind these variables
• prompts.yaml — agent prompts and persona