.env.example

GROQ_API_KEY=your-groq-key-here

# gemini is wired into primary_coder and reasoner (#72) as the mid-chain
# fallback, between Groq and ollama-colab. Unlike Groq (no card) or
# ollama-colab (keyless/local), this is a real billable Google Cloud
# credential once quota is exceeded -- set a billing/quota cap on the
# project backing this key before relying on the fallback under load. See
# models.yaml's gemini block for the rate-limit-pinning tradeoff (a
# Groq-exhausted session stays pinned on gemini's rpm_budget: 10 for the
# rest of the session rather than falling through on a 429).
GEMINI_API_KEY=your-gemini-key-here

# Optional. Each of these has a provider block in models.yaml but is wired to no
# role, so setting a key here changes nothing on its own -- point a role at the
# provider in models.yaml to actually use it.
OPENAI_API_KEY=your-openai-key-here
ANTHROPIC_API_KEY=your-anthropic-key-here

# ollama needs no key at all (api_key_required: false in models.yaml), which is
# the whole reason that flag exists -- availability used to be decided purely by
# whether a key env var was set, so a local keyless runtime was unselectable no
# matter how it was configured.

# ollama-colab is wired into primary_coder and reasoner as a last-resort
# fallback (#71). The URL is an ngrok tunnel address, session-scoped to one
# Colab run -- run colab/ollama_colab_server.ipynb (needs a free ngrok
# account + authtoken, see the notebook's first cell), copy its last cell's
# printed URL here, and update it after every Colab restart. No key needed
# for Ollama itself (api_key_required: false); the ngrok authtoken lives in
# the notebook, not here. cloudflared was tried first and dropped -- it
# consistently 403'd straight from Cloudflare's edge; see models.yaml's
# ollama-colab block for the full story.
OLLAMA_COLAB_BASE_URL=https://your-tunnel-subdomain.ngrok-free.dev/v1