Configuration

All settings are read from the environment (see .env.example). Agent-specific variables carry an AGENT_ prefix so a generic name like MODEL in your shell can't silently change the provider; API keys and LOGFIRE_TOKEN keep their standard names because the provider SDKs read those exact variables directly.

Variable	Default	Notes
Provider key (e.g. ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY, OLLAMA_BASE_URL, …)	—	Whichever variable the provider behind AGENT_MODEL reads. Not declared on Settings — Settings validates it by asking pydantic-ai to build that provider at import time, so any provider pydantic-ai supports (including ones added in later pydantic-ai releases) is checked automatically, and it raises immediately if misconfigured — not a lazy/runtime check.
AGENT_MODEL	anthropic:claude-sonnet-5-5	The agent under test. Any pydantic-ai model string works, e.g. google:gemini-2.0-flash or ollama:* for local models (no API key needed, but OLLAMA_BASE_URL must be set).
AGENT_JUDGE_MODEL	anthropic:claude-opus-5-5	Used only by the LLM-as-judge evals. Kept separate from AGENT_MODEL to avoid self-assessment bias — keep it at least as capable as the agent model, not cheaper.
LOGFIRE_TOKEN	unset	If set, traces go to Logfire cloud. If unset, traces print to the console — no separate dev-mode flag needed.
AGENT_COST_LIMIT	unset	Optional per-run spend cap in USD (e.g. 0.50). Off by default. Only set it for models with known pricing — for others (e.g. ollama:) the cost is unknown, so the cap can't be enforced and Pydantic AI warns.
AGENT_SERVICE_NAME	agent	Service name on Logfire traces. Rename it for your project.
AGENT_ENVIRONMENT	unset	Environment tag (development, production, …) on traces. Unset falls back to LOGFIRE_ENVIRONMENT.
AGENT_LOG_CONTENT	true	Whether traces include prompts, model outputs and tool arguments. Set false in production if they may be sensitive. Evals force it on, since ArgumentCorrectness reads tool arguments from spans.
AGENT_LOG_LEVEL	INFO	Standard Python logging level.

Variables only some examples read

An example reads these only if you add it; add_agent.py names the ones to set, and the release check sets the service addresses itself.

Variable	Example	Default	Notes
AGENT_EMBEDDING_MODEL	rag	follows the provider of AGENT_MODEL (Google or OpenAI)	The embedding model for retrieval. Anthropic has no embedding model, so with an Anthropic AGENT_MODEL set this, for example to google:gemini-embedding-001; the key for that provider must be set too.
CHROMA_URL	rag	http://127.0.0.1:8000	Where the Chroma server is. The Docker service publishes on a random free port, so set this to the address add_agent.py shows how to find.
MCP_SERVER_URL	mcp_tools	http://127.0.0.1:8000/mcp	The MCP server's URL. Set it to the published port, as for CHROMA_URL.
TEMPORAL_ADDRESS	temporal	127.0.0.1:7233	The host:port of the Temporal server. Set it to the published port, as for CHROMA_URL.