Configuration Reference¶

Wizard is configured through backend/.env (see backend/.env.example in the core repo for the full, current list with defaults). This page covers the settings you're most likely to actually need to touch — for the complete, byte-accurate list, the .env.example file itself is the source of truth; this page can drift, that file can't.

Providers and models¶

Variable	What it does
`API_PROVIDER`	The default provider — not a global switch. You can still assign a different provider per role (manager/worker/vision).
`MODEL_NAME` / `WORKER_MODEL_NAME` / `VISION_MODEL_NAME`	Empty by default, meaning "use whatever this provider has installed." Setting one pins it.
`DATA_MODE`	`local-only` / `cloud-only` / `hybrid`. Empty means "derive it": `local-only` on a fresh install, `cloud-only` if `API_PROVIDER` is already a cloud backend.
`DATA_SCHEMA_ONLY`	Defaults on — the conservative redaction option needs no explicit decision.
`OLLAMA_BASE_URL` / `LMSTUDIO_BASE_URL` / `OPENAI_BASE_URL` / `ANTHROPIC_BASE_URL` / `GATEWAY_API_URL`	Provider endpoints. Rewritten from `host.docker.internal` to `127.0.0.1` automatically when the backend isn't containerized, unless you've set them yourself.

Execution and sandboxing¶

Variable	What it does
`EXECUTION_BACKEND`	`host` (default, subprocess per session) / `docker` (container per session) / `inprocess` (dev/test only). `auto` and `local` are older spellings, folded to `host`.
`HOST_SANDBOX`	`off` / `best-effort` (default) / `require`. See Execution & Sandboxing.
`HOST_SANDBOX_NETWORK`	`deny` (default) / `allow` — outbound only; loopback always works.
`SANDBOX_TIER`	`core` / `standard` (default) / `full` — how much of the analysis toolkit the Docker image ships.
`SANDBOX_ENABLED`	`false` disables container creation entirely.

Agent behavior¶

Variable	What it does
`AGENT_TIER`	`auto` (default, inferred from the manager model's parameter count) / `compact` / `balanced` / `full`.
`AGENT_MAX_ITERATIONS`	A hard ceiling above whatever the tier allows — deliberately not auto-derived, since a runaway loop against a paid gateway is a billing incident.
`AGENT_REQUIRE_APPROVAL`	Turn plan-approval on. Off by default.
`AGENT_TURN_TIMEOUT`	Wall-clock deadline per turn.
`MAX_TOKENS`	The ceiling every per-purpose output budget is clamped to — lower this if you're tight on context, not the other way around.

Host sizing¶

Variable	What it does
`SYSTEM_PROFILE`	`auto` (default) measures your machine at boot and derives thread counts, worker counts, and memory limits from it.
`LLM_NUM_THREAD` / `LLM_NUM_CTX`	Left unset in the shipped example on purpose — setting either in a copied `.env` defeats the auto-derivation on every machine that copies it.

Skills¶

Variable	What it does
`SKILLS_BUILTIN_DIR` / `SKILLS_PROJECT_DIR`	Empty by default, meaning "derive it" (the checkout's `backend/skills/` and `.wizard/skills`). The user layer is never configurable — always your platform config directory, so one machine's layout can't leak into `.env.example` and confuse another.
`SKILLS_REGISTRY_API`	Left commented by default (`api.github.com` is correct for most installs) — set it for GitHub Enterprise.

What's deliberately not here¶

Repo/API secrets (model provider API keys, connector credentials) are resolved from the environment first, then a local credential store — never logged, never returned by any route. See Data Modes & Privacy.