Claude (Anthropic) — default

Scout's default provider, on the trusted Claude Agent SDK engine. The credential detection order, the zero-config Claude Code happy path, and the AI SDK caveat.

Set in scout.config.json: "model": "claude-sonnet-4-6" (any claude… id). This is Scout’s default and runs on the trusted Claude Agent SDK engine.

Credentials (detection order, first match wins)

ANTHROPIC_API_KEY env var.
CLAUDE_CODE_OAUTH_TOKEN env var.
~/.claude/.credentials.json — a real Claude Code credential file (Linux).
macOS login keychain — a Claude Code session (security entry Claude Code-credentials). Probed for existence only; never read, never prompts to unlock.

Steps 3–4 are what make the zero-config Claude Code happy path work: if you’ve run claude and signed in, Scout reuses that session with nothing to set.

Setup

Happy path — Claude Code:

# install Claude Code (https://claude.com/claude-code), then:
claude            # sign in once
scout doctor      # → ✓ credentials valid

Or an API key:

export ANTHROPIC_API_KEY=sk-ant-...
scout doctor

Notes

On macOS the Claude Code token lives in the keychain, so ~/.claude/.credentials.json is often absent even when you’re signed in — the keychain probe covers that. scout doctor may show “Detection: ✗ no credential file” and still pass on the live ping; that’s expected.
If you force the AI SDK engine (SCOUT_ENGINE=ai-sdk), the keychain/OAuth session doesn’t apply — set ANTHROPIC_API_KEY, or switch back with unset SCOUT_ENGINE.