Claude Code

Curupira runs Claude Code in print mode with streaming JSON output:

claude -p --output-format stream-json --verbose [options] -- <prompt>

Install and sign in

npm install -g @anthropic-ai/claude-code
claude   # sign in once, interactively

Claude Code owns authentication (an Anthropic account or ANTHROPIC_API_KEY). Curupira never reads or stores credentials.

Profile options

Option	Native argument	Notes
model	--model <model>	An alias such as sonnet or a full model id.
agent	--agent <agent>	A custom agent defined in Claude Code.
effort	--effort <level>	Reasoning effort level.
permission_mode	--permission-mode <mode>	For example acceptEdits or plan. Unset keeps Claude Code's default.
permission_prompts	--permission-prompts <value>	Passed through unchanged.

Unset options add no argument. The prompt always follows --, so a prompt that starts with - is not read as a flag.

Example

[agents.profiles.claude]
provider = "claude"
model = "sonnet"
permission_mode = "acceptEdits"

Sessions

The session id from the JSON stream is saved with the task. An interrupted task resumes with --resume <session_id>.