configuration.md

Configuration

CLI Parameters

Parameter	Default	Description
--directive	(required)	Natural language task description
--playbook	None	Playbook ID to execute deterministically
--playbook-params	None	JSON dict of playbook parameters
--credentials	None	JSON credentials: '{"username": "...", "password": "..."}'
--model	openai-responses:gpt-5.4	Model for LLM agent (any PydanticAI-supported model)
--max-steps	50	Max tool-call iterations (LLM path only)
--thinking	high	Thinking effort level (minimal, low, medium, high, xhigh)
--allow-private-networks	false	Allow localhost and private IPs
--start-url	None	URL to open on browser launch
--display	:99	X display (Linux)

Model Configuration

CUA uses PydanticAI and works with any model it supports — Anthropic, OpenAI, Google Gemini, Groq, and more. Set the model in settings.py:

PRIMARY_MODEL = "openai-responses:gpt-5.4"      # main agent
UTILITY_MODEL = "openai-responses:gpt-5.4-mini"  # classification, guardrails, extraction

To switch providers, change the model string and set the corresponding API key:

export ANTHROPIC_API_KEY=sk-ant-...   # Anthropic
export OPENAI_API_KEY=sk-...          # OpenAI
export GOOGLE_API_KEY=...             # Google Gemini

PydanticAI reads these automatically — no code changes needed. See PydanticAI models documentation for supported providers and model string formats.

Recommended Models

Model	Thinking	Notes
google-gla:gemini-3-flash-preview	>= high	Fastest response times. Sonnet or GPT-5.4 have better precision for complex selectors.
anthropic:claude-sonnet-4-6	>= medium	Strong selector accuracy and DOM reasoning.
openai-responses:gpt-5.4	>= medium	Strong quality. Must use openai-responses: prefix (not openai:) for thinking + tools support.

The agent relies on accurate CSS selector generation from DOM snapshots — weaker models produce selectors that timeout or miss elements.

For UTILITY_MODEL (classification, guardrails), a fast/cheap model like anthropic:claude-haiku-4-5 or openai:gpt-5.4-mini works well.