Skip to content

Configuration without the wall of settings

The demo needs no configuration:

uvx yagami demo

The base install is intentionally small. Add only the capabilities in use:

python -m pip install "yagami[providers]"  # Anthropic/OpenAI-compatible SDKs
python -m pip install "yagami[ingest]"     # PDF extraction
python -m pip install "yagami[desktop]"    # OS keyring integration
python -m pip install "yagami[all]"        # previous batteries-included behavior

For a real local setup, install and start Ollama, then initialize Yagami. The starter configuration already points to the default loopback endpoint and model:

yagami init
ollama pull llama3.2:3b-instruct-q4_K_M
yagami doctor
yagami serve

Minimum remote gateway settings

A non-loopback production bind needs headless mode plus API-key or OIDC authentication. The smallest API-key setup is:

YAGAMI_HEADLESS=true
YAGAMI_REQUIRE_AUTH=true
YAGAMI_API_KEYS=platform:replace-with-at-least-16-random-characters

Then start with yagami serve --host 0.0.0.0 --allow-remote behind a trusted TLS reverse proxy. OIDC-only authentication is also supported; configure both YAGAMI_OIDC_ISSUER and YAGAMI_OIDC_JWKS_URL instead of API keys.

Common choices

  • Local model URL/model/trust zone: YAGAMI_OLLAMA_URL, YAGAMI_OLLAMA_MODEL, YAGAMI_OLLAMA_TRUST_ZONE.
  • Cloud credentials: the provider's API-key variable, preferably injected from a secret manager or stored with Yagami's key command.
  • Production storage: YAGAMI_DATABASE_URL for PostgreSQL and YAGAMI_COORDINATION_URL for distributed coordination.
  • Signed policy enforcement: YAGAMI_POLICY_BUNDLE_PATH, YAGAMI_POLICY_PUBLIC_KEY_PATH, and YAGAMI_POLICY_SIGNATURE_REQUIRED=true.

Advanced settings

The annotated .env.example groups identity, database, policy signing, encryption, audit delivery, approvals, and external detector settings. The deployment guide explains production boundaries and the integration guide covers backend-specific settings.

Environment variables override config/yagami.toml. After changing either, restart Yagami and run yagami doctor so endpoint, model, storage, and credential problems are visible before traffic reaches the gateway.