Scout

Providers & credentials

The first run and any heal need an AI provider; deterministic replay never does. How Scout infers the provider from your model, the two engines, and how scout doctor verifies credentials.

The first scout go on a scenario (and any heal) needs an AI provider. Deterministic replay never uses an LLM, so CI that only replays needs no credentials.

Run scout doctor anytime to check: it prints the model, inferred provider, and engine, runs a fast network-free detection ladder, then does a live one-step ping to confirm the credentials actually work (exit 0 valid, 1 not — with copy-pasteable remediation).

Choosing a provider

Scout infers the provider from the model in scout.config.json (case-insensitive):

model starts withProviderGuide
claude…Anthropic (default)Claude
gemini… / google…GoogleGemini
gpt…, o1, o3, o4…OpenAIOpenAI

Unknown/empty model ids fall back to Anthropic.

Engines

  • agent-sdk (default for Anthropic) — the trusted Claude Agent SDK. Reuses a Claude Code session with zero config.
  • ai-sdk (default for Google/OpenAI; can also run Claude for parity) — the Vercel AI SDK engine. Authenticates with a real API key only.

Override with SCOUT_ENGINE=agent-sdk|ai-sdk or "engine" in scout.config.json (see the config file reference). Note: the AI SDK engine cannot reuse the Claude Code keychain/OAuth session — under it, Anthropic needs ANTHROPIC_API_KEY.

TL;DR

  • Claude Code signed in? Nothing to do — scout doctor should report ✓.
  • Otherwise export one of: ANTHROPIC_API_KEY, GEMINI_API_KEY, or OPENAI_API_KEY, matching your model.