Environment Variables

NimbleBrain reads environment variables for model provider keys, server binding, public-origin resolution, rate limiting, MCP session limits, the Composio aggregator, and observability. This page lists every supported variable with its purpose, default, and precedence.

Note

Authentication is not configured through an environment variable. Identity comes from the instance.json auth adapter (dev / oidc / workos), and the server does not start without that file. See instance.json for the full auth model.

Variable reference

Model providers

A provider’s API key can be supplied here or in nimblebrain.json under providers.<name>.apiKey. When the config field is omitted, the AI SDK falls back to the standard provider env var below.

Variable	Required	Default	Description
ANTHROPIC_API_KEY	For Anthropic models	—	Anthropic API key. Also settable via providers.anthropic.apiKey.
OPENAI_API_KEY	For OpenAI models	—	OpenAI API key. Also settable via providers.openai.apiKey.
GOOGLE_GENERATIVE_AI_API_KEY	For Google models	—	Google Gemini API key (the AI SDK’s standard Google var). Also settable via providers.google.apiKey.
NEBIUS_API_KEY	For Nebius models	—	Nebius Token Factory API key. Also settable via providers.nebius.apiKey.
XAI_API_KEY	For xAI models	—	xAI (Grok) API key. Also settable via providers.xai.apiKey.

Runtime & server

Variable	Required	Default	Description
NB_WORK_DIR	No	~/.nimblebrain	Working directory for runtime state (conversations, logs, workspaces). Does not drive config-file discovery — see config resolution.
PORT	No	27247	HTTP server port. Wins over the --port CLI flag when both are set.
ALLOWED_ORIGINS	No	—	Comma-separated additional CORS origins. The canonical origins (platform host, custom domain) are already trusted; this is for extra origins.
NB_CHAT_RATE_LIMIT	No	20	Chat requests per minute per identity.
NB_TOOL_RATE_LIMIT	No	600	Tool calls per minute per identity.
NB_MCP_RATE_LIMIT	No	300	/mcp JSON-RPC requests per minute per identity.
NB_TIMEZONE	No	Pacific/Honolulu	Default IANA timezone for time-aware features (e.g. task schedules) when an identity has no timezone preference.
NB_MAX_PARALLEL_TOOL_CALLS_PER_SOURCE	No	6	Most tool calls the agent dispatches to any one source at a time. A turn touching several connectors still fans out across them — only the depth against a single source is capped — so this bounds how loud the runtime can be toward one server, not overall throughput. Raise it for fast local servers; lower it for a fragile upstream. Zero, negative, fractional-below-1, and unparseable values fall back to the default: removing the bound is never the intended reading. Read once at startup. Scoped per source per agent run, which is a ceiling rather than a guarantee: concurrent runs each get their own budget, so the worst case against one server is roughly concurrent runs x this value, not this value. Size it with that multiple in mind.
NB_CONNECTOR_START_CONCURRENCY	No	4	How many connectors the boot loop starts in parallel, across every workspace. The default keeps peak memory and CPU bounded on a small pod while capturing most of the serial→parallel win; raise it when a deployment has many fast connectors and boot time matters, set it to 1 for a sequential start. Values below 1 and unparseable values fall back to the default. Read once at startup.
NB_CONNECTION_REVALIDATE_INTERVAL_SECONDS	No	300	Sweep interval for the connection-revalidation monitor, which re-checks connectors whose upstream credential can lapse without a transport 401 (e.g. Composio). Provider-agnostic. Blank, non-positive, or unparseable values fall back to the default. Capped at 1789569 (~20.7 days), the largest interval the runtime can schedule; a higher value logs a warning and uses the cap. To stop the sweep entirely set connectors.providers.composio.monitorEnabled: false rather than a very large interval.

When one of the chat, tool, or MCP request limits is exceeded, the API returns 429 with a Retry-After header containing the seconds remaining in that caller’s fixed one-minute window, rounded up.

Public origin and callbacks

The platform builds every outward-facing URL (OAuth callbacks, post-login landing) from a single canonical public origin. Resolution order: NB_PUBLIC_ORIGIN → derived from NB_PLATFORM_HOST / NB_CUSTOM_DOMAIN → http://localhost:27247 (dev).

Variable	Required	Default	Description
NB_PUBLIC_ORIGIN	Recommended (production)	derived, else http://localhost:27247	Canonical public origin (scheme + host, no path). The explicit operator override — trusted as given. Used to build the OAuth callback ${NB_PUBLIC_ORIGIN}/v1/mcp-auth/callback and the WorkOS / Composio callbacks. Must be a bare https origin (or http on a loopback host in dev); a value with a path/query throws at startup.
NB_PLATFORM_HOST	No	—	Platform subdomain host (no scheme). Used to derive the canonical origin when NB_PUBLIC_ORIGIN is unset. Chart-forwarded in production.
NB_CUSTOM_DOMAIN	No	—	Custom domain host. When set and canonical, becomes the derived public origin (custom domain wins).
NB_CUSTOM_DOMAIN_CANONICAL	No	true	"true" / "false" only — anything else throws. When "false", the custom domain is served but the platform host stays canonical.
NB_WEB_URL	No	falls back to the canonical origin	User-facing SPA origin for post-OAuth browser returns. Only needed when the API and SPA are served from different origins (e.g. dev: API on :27247, SPA on :27246).

Connectors catalog

Variable	Required	Default	Description
NB_CURATED_CATALOG_DIR	No	in-image example	The connector catalog — a directory of ServerDetail YAML/JSON files (the intended use, hence the name), or a single such file. Every *.yaml/*.yml/*.json in it is read in sorted filename order and re-read on each lookup, so editing a mounted ConfigMap takes effect without a restart. Replaces src/connectors/catalog/curated/ wholesale. This is where auth: static and brokered (composio, smithery) entries belong; the in-image example is DCR-only. See Connectors Catalog.

Secrets

Variable	Required	Default	Description
WORKOS_API_KEY	When auth.adapter: "workos" and no apiKey in instance.json	—	WorkOS API key. Also settable via instance.json → auth.apiKey.

Caution

NB_API_KEY is not an authentication gate. It does not enable, configure, or bypass auth. To configure auth, use instance.json.

Composio aggregator

Required only when the connectors catalog contains any auth: composio entries (Gmail, Outlook, …). See the Composio operator guide for end-to-end setup.

The provider’s settings live in nimblebrain.json under connectors.providers.composio.

COMPOSIO_API_KEY is the one value with two sources — it is also declarable as connectors.providers.composio.apiKey, though the environment remains the better posture for a secret (why). NB_TENANT_ID is env-only.

Variable	Required	Default	Description
COMPOSIO_API_KEY	When the catalog has auth: composio entries	—	(fallback for …composio.apiKey) Platform-wide Composio API key — the broker credential, and the gate on registering the provider at all. Read once at startup. Without it, the Composio integration reports “not configured” and installs of auth: composio entries fail.
NB_TENANT_ID	When multi-tenant pods share a Composio account	—	Per-pod tenant identifier prepended to the wsId so two tenants never collide in Composio’s namespace. Required in bouncer (multi-tenant) mode — startup fails without it. Single-tenant deployments can omit it.

Smithery broker

Required only when the connectors catalog contains any auth: smithery entries. Smithery brokers the connection and hosts the MCP session; the platform stores no upstream credential.

A credential and a namespace must both resolve — a Smithery namespace is globally unique and account-owned, so there is no safe default. Setting only the key logs a warning and leaves the integration disabled.

SMITHERY_API_KEY is a fallback for connectors.providers.smithery.apiKey, because an installed connector’s transport names the smithery credential provider rather than an env var. Every other Smithery setting lives in that block only.

Variable	Required	Default	Description
SMITHERY_API_KEY	When the catalog has auth: smithery entries	—	Platform-wide Smithery API key, from smithery.ai/account/api-keys. Read once at startup. Fallback for connectors.providers.smithery.apiKey.
NB_TENANT_ID	When multi-tenant pods share a Smithery account	—	Prepended to the workspace id so two tenants never collide in the namespace. Same variable as the Composio prefix above.

Observability (OpenTelemetry + logs)

Vendor-neutral. Trace ids always exist for log correlation; export only happens when an OTLP endpoint is configured, so local dev and OSS checkouts need no infrastructure.

Variable	Required	Default	Description
OTEL_EXPORTER_OTLP_ENDPOINT	No	— (no export)	OTLP collector base URL. Unset means spans are still created (for log correlation) but nothing is exported.
NB_SERVICE_NAME	No	—	Overrides the service name stamped on spans and structured logs.
NB_TENANT_ID	No	—	Boot-time resource attribute stamped on spans/logs as tenant_id. (Same var as the Composio multi-tenant prefix above.)
NB_LOG_FORMAT	No	pretty	Set to json for structured JSON logs auto-enriched with service, tenant_id, correlation_id (the active trace id), and identity. Set by the chart in production.
NB_DEBUG	No	—	Namespace-scoped hot-path debug logging (e.g. NB_DEBUG=mcp,sse). Read once at process start.

Error reporting (Sentry)

Optional, additive error capture on top of OpenTelemetry. Off unless NB_SENTRY_ENABLED=true (the SDK is loaded only then — never inferred from DSN presence), initialized as an out-of-kernel preload that does not take over OTel tracing. NB_SENTRY_* are our own knobs (mirroring the web client), not Sentry’s auto-read SENTRY_*. Per-tenant via the chart (runtime.config.sentry.enabled). See Observability → Error reporting.

Variable	Required	Default	Description
NB_SENTRY_ENABLED	No	false	Explicit on/off switch. Must be true to enable; never inferred from the DSN.
NB_SENTRY_DSN	No	—	Project DSN (required when enabled).
NB_SENTRY_ENV	No	—	Environment tag (e.g. staging, production); one project discriminated by this tag.
NB_SENTRY_TRACES_SAMPLE_RATE	No	0	Sentry performance tracing. 0 = errors only (recommended; leaves tracing to OTel).

Release is taken from the runtime’s existing NB_VERSION → NB_BUILD_SHA (no separate Sentry release var).

Telemetry

Variable	Required	Default	Description
NB_TELEMETRY_DISABLED	No	—	Set to 1 to disable anonymous usage telemetry.
DO_NOT_TRACK	No	—	Standard opt-out signal. Set to 1 to disable telemetry. Equivalent to NB_TELEMETRY_DISABLED=1.
POSTHOG_API_KEY	No	—	Override the built-in PostHog key. Rarely needed.

Web client (runtime config)

The web UI is a static build, so it cannot read pod env from JS. Instead Caddy serves /config.js dynamically from these NB_* values (set window.__NB_CONFIG__), which the client reads on boot. This lets one image be configured per tenant via Helm with no rebuild and no writable filesystem. All values are public client keys — never secrets. Unset ⇒ the corresponding feature is disabled (the OSS/self-host default).

Variable	Required	Default	Description
NB_SENTRY_DSN	No	—	Sentry DSN (public). Unset ⇒ client crash tracking off.
NB_SENTRY_ENABLED	No	false	Set true to enable Sentry for this deployment (also requires a DSN).
NB_SENTRY_ENV	No	—	Sentry environment label (e.g. staging, production).
NB_SENTRY_TRACES_SAMPLE_RATE	No	0	Performance-trace sample rate, 0–1.
NB_POSTHOG_KEY	No	—	PostHog write-only key for web analytics.
NB_POSTHOG_ENABLED	No	false	Set true to enable web analytics (also requires a key).

Sentry tags every event with tenant_id (from NB_TENANT_ID), workspace_id, and an opaque user.id — never email, display name, prompts, or tool content. For local development, the same values can be supplied to bun run dev as VITE_* fallbacks (see .env.example).

NB_PUBLIC_ORIGIN

The canonical public origin of the deployment. Every OAuth callback URL the platform hands to a vendor is built from it:

${NB_PUBLIC_ORIGIN}/v1/mcp-auth/callback

Each OAuth app registered with a vendor (Asana, HubSpot, Gmail, …) must list this exact URL in its redirect_uri allow-list. A mismatch surfaces as a vendor-side redirect_uri does not match error at the moment a user clicks Connect.

export NB_PUBLIC_ORIGIN=https://nimblebrain.acme.com

The value must be a bare origin — scheme + host with no path, query, or fragment. A malformed value (a pasted full callback URL, an http origin on a non-loopback host) throws at startup rather than minting broken callbacks at first click.

In dev (bun run dev) the default http://localhost:27247 is correct without setting the variable. In production, set this explicitly (or supply NB_PLATFORM_HOST / NB_CUSTOM_DOMAIN for the chart to derive it). Pointing it at localhost, an internal cluster service name, or a stale staging URL produces confusing OAuth failures that look like vendor outages.

Tip

The Set up modal in the UI shows the resolved redirect URI with a Copy button. Paste that exact string into the vendor portal.

NB_WORK_DIR

Sets the working directory where NimbleBrain stores runtime state.

export NB_WORK_DIR=/data/nimblebrain

The working directory contains workspaces/<wsId>/ (which holds each workspace’s conversations, files, and tasks), users/<userId>/, logs/, and cache/.

NB_WORK_DIR sets the workDir runtime value (where state lives). It does not select which nimblebrain.json is loaded — config-file discovery is driven by --config, a project-local .nimblebrain/, and the command’s default workdir. See config resolution.

ALLOWED_ORIGINS

Adds origins to the CORS allowlist beyond the canonical ones (platform host and custom domain, which are trusted automatically).

export ALLOWED_ORIGINS=https://app.example.com,https://admin.example.com

Separate multiple origins with commas. Whitespace around each origin is trimmed.

Telemetry variables

Either variable disables anonymous telemetry:

export NB_TELEMETRY_DISABLED=1
export DO_NOT_TRACK=1

DO_NOT_TRACK follows the Console Do Not Track standard. You can also disable telemetry via telemetry.enabled: false in nimblebrain.json.

Tip

Environment variables are checked before the config file, so NB_TELEMETRY_DISABLED or DO_NOT_TRACK turn telemetry off regardless of what nimblebrain.json says.

Precedence summary

When the same setting can be configured in multiple places, this is the resolution order (highest priority first):

Setting	Env var	CLI flag	Config file	Default
Working directory	NB_WORK_DIR (beats all)	--workdir	workDir	~/.nimblebrain
Port	PORT (beats --port)	--port	http.port	27247
Public origin	NB_PUBLIC_ORIGIN	—	derived from host facts	http://localhost:27247
Anthropic key	ANTHROPIC_API_KEY	—	providers.anthropic.apiKey	—
Telemetry	NB_TELEMETRY_DISABLED / DO_NOT_TRACK	—	telemetry.enabled	Enabled

Example: production environment

A typical production setup:

export ANTHROPIC_API_KEY=sk-ant-api03-...
export NB_PUBLIC_ORIGIN=https://app.example.com
export NB_WORK_DIR=/data/nimblebrain
export NB_TELEMETRY_DISABLED=1
# Composio connectors (only if the catalog uses auth: composio).
# Per-toolkit ac_… ids are not env vars — declare them in nimblebrain.json
# under connectors.providers.composio.authConfigs.
export COMPOSIO_API_KEY=...
# Observability
export NB_LOG_FORMAT=json
export OTEL_EXPORTER_OTLP_ENDPOINT=http://otel-collector:4318

Authentication is configured separately in instance.json, not via environment variables.