Configuration (env)

Copy .env.example → .env. The only required variable is an LLM key.

Variable	Default	Purpose
LLM_PROFILE	anthropic	anthropic | openrouter | mixed — which provider profile (per-tier default models)
LLM_ROUTING	—	Per-role preset: fast (Groq worker) | volume (OpenRouter worker) — see Role routing
ANTHROPIC_API_KEY	—	API key for the Anthropic profile
OPENROUTER_API_KEY	—	API key for the OpenRouter profile (cheaper)
GROQ_API_KEY	—	API key for the Groq worker (used by LLM_ROUTING=fast)
QA_TESTCASE_LANG	English	Language for generated test cases, e.g. Ukrainian / uk
BROWSER_BACKEND	lib	lib (in-process Playwright) | cli — how the browser is driven
BROWSER_CHANNEL	—	chrome to use real Chrome (helps with OAuth)
MAX_REPAIR	2	Max self-repair attempts per failing spec (keep-best)
LANGFUSE_BASE_URL	—	Langfuse — cloud or self-hosted (optional)
LANGFUSE_PUBLIC_KEY	—	Langfuse public key
LANGFUSE_SECRET_KEY	—	Langfuse secret key

Tip

Every variable above can also be set with a CAIRN_ prefix (e.g. CAIRN_LLM_PROFILE, CAIRN_MAX_REPAIR) — handy for namespacing Cairn’s config alongside other tools.

Note

Langfuse is all-or-nothing: set all three LANGFUSE_* variables or none. With none set the bot runs fully offline — see Langfuse (optional).

Minimal .env

ANTHROPIC_API_KEY=sk-ant-...

Role routing

By default each profile uses its own per-tier models. LLM_ROUTING optionally layers a cheap worker over any profile while keeping the strong reasoner — one flag picks where the mechanical steps (identify elements, generate code, repair) run:

LLM_ROUTING	Worker	Notes
unset (default)	the profile’s own per-tier models	no routing
fast	Groq llama-3.3-70b-versatile	lowest latency / cost (needs GROQ_API_KEY)
volume	OpenRouter deepseek/deepseek-chat	model breadth (needs OPENROUTER_API_KEY)

In every preset the reasoner — designing test cases and the Pilot verdict — stays on Anthropic claude-opus-4-8 for judgment quality. Override any role with CAIRN_ROLE_WORKER / CAIRN_ROLE_REASONER=provider:model, or pass --routing <preset> on explore / design / automate to set it per run. Per-run, per-role cost (tokens + $) is printed in the run summary — see ADR 0011.

Going further

• Steer the design pass with domain knowledge files & prompt overrides
• Saved sessions live in .auth/<name>.storageState.json — capture them with cairn session capture --url <url> --name <name>