Configuration

Environment variables, the yaml config file, and model provider options.

Environment variables

Variable	Required	Description
`ANTHROPIC_API_KEY`	Yes (Claude default)	Anthropic API key
`ANTHROPIC_BASE_URL`	No	API endpoint (default: `https://api.anthropic.com`)
`AGENT_PROVIDER`	No	`claude` (default) or `codex`
`CODEX_API_KEY`	Yes (if Codex)	OpenAI Codex API key
`CODEX_BASE_URL`	No	Custom Codex API base URL
`SAST_INTERNAL_TOOLS_MCP_URL`	No	MCP server URL (alternative to `--mcp-server-url`)
`SAST_INTERNAL_TOOLS_MCP_BEARER`	No	Bearer token for MCP HTTP auth

CLI flags override env where noted (e.g. --provider, --mcp-server-url).

Configuration file

Default path: conf/appsec_agent.yaml (bundled with the package).

Structure:

default:
  threat_modeler:
    options:
      max_turns: 100
  pr_reviewer:
    options:
      diff_review_max_tokens_per_batch: 150000
      diff_review_max_batches: 3

environment — top-level key (default: default); pass -e / --environment to select
options.max_turns — tool-use turn limit per role
pr_reviewer.options — PR chunking settings (see Examples — PR chunking)

Override the file path: -y /path/to/appsec_agent.yaml

Override turns for any role: --max-turns <n>

Model providers

Since v3.0.0, every role uses a provider-neutral RoleSpec. Pick the backend at runtime.

Claude (default)

Uses the Anthropic API via @anthropic-ai/claude-agent-sdk.

npx agent-run -r code_reviewer -s ./src -m sonnet

Model aliases: sonnet, opus, haiku, or full IDs like claude-sonnet-4-6.

Codex (opt-in)

Uses @openai/codex-sdk. Requires CODEX_API_KEY (or --provider codex with key in env).

export CODEX_API_KEY="..."
npx agent-run -r pr_reviewer --diff-context pr.json -s ./repo \
  --provider codex -m gpt-4.1 -f json

Model IDs: gpt-*, o* (e.g. gpt-4.1, o3). Claude aliases are mapped automatically (opus → o3, sonnet → gpt-4.1).

Set globally:

export AGENT_PROVIDER=codex

MCP on both providers

When --mcp-server-url is set, supported roles attach an HTTP MCP server exposing:

queryFindingsHistory
queryImportGraph
queryRuntimeEnrichment
queryCodebaseGraph

MCP-aware roles: pr_reviewer, code_reviewer, pr_adversary, fp_adversary, finding_validator, code_fixer.

Override server name (affects tool prefix): --mcp-server-name my-server

Default name: appsec-internal → tools appear as mcp__appsec-internal__queryFindingsHistory, etc.

Parent apps can pass URL via env instead of argv:

export SAST_INTERNAL_TOOLS_MCP_URL="http://127.0.0.1:9999/mcp"
export SAST_INTERNAL_TOOLS_MCP_BEARER="token-if-needed"

CLI security notes

API keys: Use env vars, not -k/--anthropic-api-key, in production
Input files: Paths are validated against directory traversal before reading
Output files: Validated similarly on write

Per-role output formats

Role	Default output	Structured schema
`code_reviewer` / `pr_reviewer`	markdown	`security_review_report` (JSON with `-f json`)
`pr_adversary`	JSON only	`security_review_report`
`fp_adversary`	JSON only	`fp_adversary_report`
`threat_modeler`	markdown	`threat_model_report` (JSON with `-f json`)
`code_fixer`	JSON	`FixOutput`
`qa_verifier`	JSON	`QaVerdict`

Schema sources live in src/schemas/.