Configuration Guide

For provider-specific credential and endpoint examples, see Provider Configuration.

Environment Variables

Create a .env file from the template:

cp .env.example .env

Variable	Description	Default
API_KEY	API key (supports OpenAI and compatible providers)	Required for openai provider
OPENAI_BASE_URL	Custom API endpoint	https://api.openai.com/v1
OPENAI_MODEL	Default model name in .env.example	gpt-4o
LLM_PROVIDER	Default provider in .env.example	openai
ANTHROPIC_API_KEY	Anthropic API key	Required for anthropic provider
DEFAULT_DEFENSE	Default defense strategy	D0
MAX_AGENT_STEPS	Max reasoning steps	10
AGENT_TEMPERATURE	LLM temperature	0.7

YAML Configuration

config/tools.yaml

Defines available tools with risk metadata:

tools:
  - name: send_email
    risk_level: high
    side_effect: true
    parameters:
      to:
        whitelist: ["user@company.com", "team@company.com"]

config/defenses.yaml

Configures defense strategy parameters:

defenses:
  D1:
    config:
      delimiter_start: "<<UNTRUSTED>>"
      add_warning: true
  D2:
    config:
      block_critical_tools: true
      high_risk_tools: [send_email, write_file]

config/models.yaml

Defines LLM models with pricing information for cost tracking.

Third-Party API Endpoints

To use OpenAI-compatible endpoints (vLLM, Ollama, LiteLLM, etc.):

# Via CLI
asb run "task" --provider openai-compatible --base-url https://your-proxy.com/v1 --model your-model

# Via environment
export OPENAI_BASE_URL=https://your-proxy.com/v1
export LLM_PROVIDER=openai-compatible