Configuration

All configuration is optional. With nothing set, the app starts and explains how to finish setup. Keys are read from the environment or a gitignored .env.

Environment variables

Variable	Default	Purpose
OLLAMA_BASE_URL	http://localhost:11434	Local model endpoint
OLLAMA_MODEL	llama3.2	Local model name
GROQ_API_KEY	—	Enables the Groq free-tier fallback
GROQ_MODEL	llama-3.3-70b-versatile	Groq model
OPENROUTER_API_KEY	—	Enables the OpenRouter fallback
OPENROUTER_MODEL	meta-llama/llama-3.3-70b-instruct:free	OpenRouter model
JARVIS_DATA_DIR	OS app-data dir	Where memory, events, notes, and skills live
JARVIS_STOP	—	Kill switch — set to halt any autonomous loop

Routing order

The router always tries Ollama first (local, private, unlimited), then each configured cloud provider in order. If nothing is reachable, it returns a friendly onboarding message rather than an error.

Free-tier hygiene

• Response cache — identical requests are served from a 10-minute cache (deduping double-sends and retries). Cache hits are labeled cached in the reply footer, never served silently.
• Rate-limit backoff — a cloud provider that returns 429/503 goes on an exponential cooldown (30s → 10min, honoring Retry-After) and is skipped until it clears. The local model is exempt — it’s unlimited.

Data directory layout

<data_dir>/
├─ jarvis.sqlite3     # messages, facts, insights (WAL mode)
├─ events.jsonl       # append-only action log
├─ notes/             # <slug>.md
└─ skills/            # <slug>/{manifest.json, skill.rhai, test.rhai, history/}

The store is exportable and wipeable from the Memory tab — see Export or wipe your memory.