Environment variables

Environment variables

Important: Poindexter runs virtually ALL configuration out of the app_settings Postgres table, not environment variables. Env vars and ~/.poindexter/bootstrap.toml exist only to bootstrap the stack (enough config to connect to the database and start Docker). Once the worker connects, everything — model thresholds, prompt templates, QA weights, auto-publish rules, algorithm windows, SEO pinging endpoints — is changed with SQL against app_settings (hundreds of keys — see docs/reference/app-settings.md for the full, nightly-regenerated catalog), no redeploy needed.

Preferred path: run poindexter setup once on first install. It writes ~/.poindexter/bootstrap.toml with 5 bootstrap values (database URL + 4 generated passwords). Then bash scripts/start-stack.sh reads bootstrap.toml and exports them for Docker. No .env file needed.

See reference/app-settings.md for the full list of DB-backed settings.

This document enumerates the environment variables consumed by the stock docker-compose.local.yml deployment. If you’re running a custom configuration, start from this list.

---

Required

These variables are generated by poindexter setup and stored in ~/.poindexter/bootstrap.toml. The scripts/start-stack.sh wrapper exports them automatically for Docker. Missing any causes the stack to fail fast with a clear error — no silent fallbacks.

Variable	Purpose	Example
LOCAL_POSTGRES_PASSWORD	Password for the local Postgres superuser	random 32-char hex
GRAFANA_PASSWORD	Grafana admin password	random 32-char hex
PGADMIN_PASSWORD	pgAdmin admin password	random 32-char hex

All three are generated automatically by poindexter setup on first run and written to ~/.poindexter/bootstrap.toml (safe permissions, outside the repo). No .env file is needed.

Worker authentication uses OAuth 2.1 client credentials (Phase 3 finalised in Glad-Labs/poindexter#249). poindexter setup provisions an initial CLI client and stores the credentials encrypted in app_settings.cli_oauth_client_id / cli_oauth_client_secret. There is no API_TOKEN env var.

Optional (with sensible defaults)

Variable	Default	Purpose
API_BASE_URL	http://localhost:8002	Base URL clients use to reach the worker
COMPANY_NAME	Glad Labs	Publisher name used in generated content’s <author> tag
DEFAULT_OLLAMA_MODEL	auto	Ollama model the router picks when no override
DEVELOPMENT_MODE	true	Enables dev bypasses (token, CORS). Never set in prod.
ENVIRONMENT	production	Log tag — also tagged on GlitchTip events. Sentry SDK is back via self-hosted GlitchTip; main.py initialises it when app_settings.sentry_dsn is set (provisioned 2026-05-09).
GRAFANA_USER	admin	Grafana admin username
LOCAL_POSTGRES_USER	poindexter	Local Postgres user
LOCAL_POSTGRES_DB	poindexter_brain	Local Postgres database name
OLLAMA_BASE_URL	http://host.docker.internal:11434	Where OllamaClient reaches Ollama (pipeline generation)
OLLAMA_URL	http://host.docker.internal:11434	Where MemoryClient / auto-embed.py reaches Ollama (embeddings). Must be set to the same value as OLLAMA_BASE_URL.
PORT	8002	Worker HTTP port
SITE_DOMAIN	gladlabs.io	The operator’s public site domain
SITE_NAME	Glad Labs	Site name in feeds + metadata
SITE_URL	https://www.gladlabs.io	Canonical site URL

Optional (no default — feature off unless set)

Variable	Feature gated on it
PEXELS_API_KEY	Fallback stock-photo search when image-gen is unavailable
TELEGRAM_BOT_TOKEN	Brain daemon alerts via Telegram
TELEGRAM_CHAT_ID	Destination chat for Telegram alerts
OPENCLAW_GATEWAY_URL	Discord + Telegram bridge via OpenClaw
OPENCLAW_GATEWAY_TOKEN	Bearer token for OpenClaw Tools Invoke API
POINDEXTER_MCP_HTTP_TOOL_ALLOWLIST	Comma-separated tool names exposed by the remote MCP HTTP transport. Unset = expose all 25+ tools. Set to the recommended voice/mobile read-only list when registering a Custom Connector to a phone (see below)

Local Ollama only in the core stack. Paid-API providers (OpenAI, Anthropic, Google) live in community plugins, never the core. If you install one of those plugins, it reads its own keys from app_settings at runtime via the settings API — not from env vars.

.env.local template

A fully-populated .env.local for a standard local install looks like this (the bootstrap script writes something equivalent):

# ---- Required ----
# Note: API_TOKEN was removed in Phase 3 (Glad-Labs/poindexter#249).
# Worker auth is OAuth 2.1 only — `poindexter setup` provisions the
# initial CLI OAuth client automatically.
LOCAL_POSTGRES_PASSWORD=REDACTED
GRAFANA_PASSWORD=REDACTED
PGADMIN_PASSWORD=REDACTED

# ---- Optional overrides ----
COMPANY_NAME=Your Brand Here
SITE_NAME=Your Brand Here
SITE_DOMAIN=yourbrand.com
SITE_URL=https://www.yourbrand.com

# ---- Feature flags ----
DEVELOPMENT_MODE=true
ENVIRONMENT=production
PORT=8002

# ---- Optional integrations (leave empty if you don't use them) ----
PEXELS_API_KEY=
TELEGRAM_BOT_TOKEN=
TELEGRAM_CHAT_ID=

MCP HTTP tool allowlist (POINDEXTER_MCP_HTTP_TOOL_ALLOWLIST)

The remote MCP HTTP server (mcp-server/http_server.py) exposes the same tool registry as the stdio transport — by default that’s all 25+ tools, including write-capable ones like publish_post, approve_post, reject_post, set_setting, and create_post.

When you register the HTTP server as a Custom Connector for the mobile or voice Claude app, write tools become a real risk: a misheard query plus an LLM that “follows through” can publish or reject content from the road, exactly when you’re least able to course-correct quickly.

Set POINDEXTER_MCP_HTTP_TOOL_ALLOWLIST to the comma-separated list below to expose only the 13 read-only tools — covers “what’s the system state?”, “what did I decide about X?”, and “what’s in the queue?” without giving voice mode the keys to the publishing button:

export POINDEXTER_MCP_HTTP_TOOL_ALLOWLIST="search_memory,recall_decision,find_similar_posts,list_tasks,get_post_count,get_setting,list_settings,get_audit_log,get_audit_summary,get_brain_knowledge,check_health,get_budget,memory_stats"

The same list lives as the DEFAULT_VOICE_MOBILE_ALLOWLIST constant in mcp-server/http_server.py for in-code reference.

Behaviour notes:

• Unset ⇒ all registered tools exposed (pre-#239 default).
• Empty string ("") ⇒ explicit “expose nothing” — useful if you want the route mounted but locked down without unmounting it.
• Whitespace-tolerant — a, b ,c parses as {a, b, c}.
• Unknown names silently ignored — copying the recommended list verbatim never crashes startup, even after a tool is renamed upstream.
• Stdio transport unaffected — running python mcp-server/server.py always exposes the full registry, since the stdio path doesn’t go through http_server.build_app.

The filter runs before MCPServer builds its streamable_http_app(), so both tools/list and tools/call reflect the trimmed registry — unlisted tools surface as Unknown tool: <name> to the client.

How the worker actually reads these

The worker resolves configuration in this order:

1. Environment variable — for bootstrap-only values (DATABASE_URL, passwords). Read once at startup. Worker API auth uses OAuth JWTs (no env var); see poindexter auth migrate-cli.
2. app_settings table — for runtime-mutable values (model thresholds, prompt overrides, cost guards, feature flags). Consulted on every call via site_config.get(key, default).
3. Hardcoded fallback — the _DEFAULTS dict in services/site_config.py. Used only when the DB value is absent AND no env var provides a value.

Never edit _DEFAULTS to change runtime behavior — that requires a rebuild. Use the settings API or SQL against app_settings instead.

Where to go next

• Full settings catalog: reference/app-settings.md
• Local dev setup: local-development-setup.md
• Troubleshooting: troubleshooting.md