.env.example

# BEGIN GENERATED CONNECTIONS
# Generated by scripts/generate_env_example.py; definitions live in prometheus/integrations.py

PROMETHEUS_PROVIDER=anthropic

# ── AI Providers ──

# ── Anthropic ──
# Uses Anthropic models.
# ANTHROPIC_API_KEY=

# ── Openai ──
# Uses Openai models.
# OPENAI_API_KEY=

# ── Openrouter ──
# Uses Openrouter models.
# OPENROUTER_API_KEY=

# ── Gemini ──
# Uses Gemini models.
# GEMINI_API_KEY=

# ── Deepseek ──
# Uses Deepseek models.
# DEEPSEEK_API_KEY=

# ── Minimax ──
# Uses Minimax models.
# MINIMAX_API_KEY=
# Choices: https://api.minimaxi.com/anthropic, https://api.minimax.io/anthropic.
# MINIMAX_BASE_URL=https://api.minimaxi.com/anthropic

# ── Kimi ──
# Uses Kimi models.
# MOONSHOT_API_KEY=
# Choices: https://api.moonshot.ai/anthropic, https://api.moonshot.cn/anthropic.
# MOONSHOT_BASE_URL=https://api.moonshot.ai/anthropic

# ── Glm ──
# Uses Glm models.
# ZHIPU_API_KEY=
# Choices: https://api.z.ai/api/anthropic, https://open.bigmodel.cn/api/anthropic.
# ZHIPU_BASE_URL=https://api.z.ai/api/anthropic

# ── Xai ──
# Uses Xai models.
# XAI_API_KEY=

# ── Opencode Zen ──
# Uses Opencode Zen models.
# OPENCODE_ZEN_API_KEY=

# ── Opencode Go ──
# Uses Opencode Go models.
# OPENCODE_GO_API_KEY=

# ── Channels ──

# ── Telegram ──
# Lets Prometheus receive and send Telegram messages.
# Install: pip install -e '.[telegram]'
# Setup: https://t.me/BotFather
# TELEGRAM_BOT_TOKEN=
# TELEGRAM_ALLOWED_USER=

# ── Discord ──
# Lets Prometheus answer in a Discord server.
# Install: pip install -e '.[discord]'
# DISCORD_BOT_TOKEN=
# DISCORD_ALLOWED_USER=
# DISCORD_ALLOWED_CHANNEL=
# Set to 1 to enable.
# DISCORD_REQUIRE_MENTION=
# DISCORD_MAX_TURNS_PER_HOUR=
# DISCORD_HOME=

# ── WhatsApp ──
# Lets Prometheus answer WhatsApp messages via the Meta Cloud API.
# Install: pip install -e '.[whatsapp]'
# Setup: https://developers.facebook.com/apps
# WHATSAPP_TOKEN=
# WHATSAPP_PHONE_NUMBER_ID=
# WHATSAPP_APP_SECRET=
# WHATSAPP_VERIFY_TOKEN=
# Without '+' prefix. Empty = answer anyone.
# WHATSAPP_ALLOWED_PHONE=

# ── Calendar & Productivity ──

# ── Google Calendar ──
# Lets Prometheus create and update Google Calendar events.
# Install: pip install -e '.[gcal]'
# Setup: https://developers.google.com/calendar/api/quickstart/python
# Set to 1 to enable.
# PROMETHEUS_GOOGLE_CALENDAR=
# PROMETHEUS_GOOGLE_CALENDAR_ID=primary

# ── Apple Calendar ──
# Lets Prometheus work with Apple Calendar on this Mac.
# Set to 1 to enable.
# PROMETHEUS_APPLE_CALENDAR=
# PROMETHEUS_APPLE_CALENDARS=

# ── Apple Tools ──
# Lets Prometheus use Apple Mail and other local Apple tools.
# Set to 1 to enable.
# PROMETHEUS_APPLE_TOOLS=

# ── Memory & Storage ──

# ── Notion ──
# Stores episodic memory in a Notion database.
# Install: pip install -e '.[notion]'
# Setup: https://www.notion.so/my-integrations
# Choices: sqlite, notion.
# PROMETHEUS_EPISODIC_STORE=sqlite
# NOTION_TOKEN=
# NOTION_EPISODES_DATABASE_ID=

# ── Mem0 ──
# Stores semantic memory in the Mem0 service.
# Install: pip install -e '.[arena]'
# Choices: sqlite, mem0.
# PROMETHEUS_SEMANTIC_STORE=sqlite
# From app.mem0.ai. The adapter sends infer=False so add() always stores — a bake-off against it measures retrieval, not Mem0's own extraction step.
# MEM0_API_KEY=
# Defaults to 'prometheus'. Change it only if several people share one Mem0 account.
# MEM0_USER_ID=

# ── Zep ──
# Stores semantic memory in a Zep temporal graph.
# Install: pip install -e '.[arena]'
# Choices: sqlite, zep.
# PROMETHEUS_SEMANTIC_STORE=sqlite
# From getzep.com. Facts become graph edges with validity dates, so a correction supersedes the old value instead of ranking beside it — the one backend built for the update test.
# ZEP_API_KEY=
# Defaults to 'prometheus'. Zep scopes a graph per user.
# ZEP_USER_ID=
# Ingestion is asynchronous: graph.add returns in ~0.2s with the episode unprocessed, and the text is not searchable until Zep has turned it into nodes and edges. Prometheus polls until it has, up to this long. Default 120 — raise it if seeding times out, never lower it to make a benchmark finish.
# ZEP_MAX_WAIT_SECONDS=

# ── LangMem ──
# Stores semantic memory in a LangGraph store via LangMem.
# Install: pip install -e '.[arena]'
# Choices: sqlite, langmem.
# PROMETHEUS_SEMANTIC_STORE=sqlite
# Optional. Without it LangGraph's InMemoryStore is used, which its own docs describe as a reference implementation whose data dies with the process — fine for a benchmark run, not a persistence story.
# PROMETHEUS_LANGMEM_POSTGRES=
# LangMem has no key of its own; it bills through embeddings. Semantic search needs this or the store is a plain key-value dict.
# OPENAI_API_KEY=

# ── Supabase ──
# Stores semantic memory in Supabase pgvector.
# Install: pip install -e '.[supabase]'
# Choices: sqlite, supabase.
# PROMETHEUS_SEMANTIC_STORE=sqlite
# SUPABASE_URL=
# Embeddings also require OPENAI_API_KEY (optionally OPENAI_EMBED_MODEL).
# SUPABASE_SERVICE_KEY=

# ── Search & Observability ──

# ── Tavily ──
# Lets Prometheus search the web.
# Setup: https://tavily.com
# TAVILY_API_KEY=

# ── treg ──
# Your own treg key. treg then bills your treg account, not your Prometheus credits.
# Setup: https://treg.to
# An org-scoped key from your treg dashboard. It is sent only to treg.
# TREG_API_KEY=

# ── Memory & Storage ──

# ── TypeSafe Jev ──
# Lets Jev decide which memories earn a place in each answer.
# Setup: https://typesafe.ai
# TYPESAFE_API_KEY=
# PROMETHEUS_SLOT_GATE=

# ── Search & Observability ──

# ── OpenTelemetry ──
# Exports traces to an OTLP collector.
# Install: pip install -e '.[tracing]'
# OTEL_EXPORTER_OTLP_ENDPOINT=

# END GENERATED CONNECTIONS

# ── Models (optional) ───────────────────────────────────────────────────────
# Each provider has sensible defaults (see prometheus/loop/models.py PROVIDERS):
#   anthropic:  claude-sonnet-5              + claude-haiku-4-5 (gate/summarizer)
#   openai:     gpt-5.5                      + gpt-4.1-mini
#   gemini:     gemini-3.5-flash             + gemini-3.1-flash-lite
#   deepseek:   deepseek-v4-pro              + deepseek-v4-pro
#   minimax:    MiniMax-M3                   + MiniMax-M2
#   kimi:       kimi-k3                      + kimi-k2.6 (gate/summarizer)
#               (K3-only key? set PROMETHEUS_SMALL_MODEL=kimi-k3 too)
#   glm:        glm-5.2                      + glm-5-turbo
#   openrouter: nemotron-3-super:free        + gemma-4-26b:free (both $0,
#               rate-limited; Settings lists the live catalog, free first)
# openrouter's whole point is picking a different model, so PROMETHEUS_MODEL takes
# any "<vendor>/<model>" slug from https://openrouter.ai/models — e.g.
# PROMETHEUS_MODEL=meta-llama/llama-3.1-70b-instruct
# Override here if you want something else — they're just strings:
# PROMETHEUS_MODEL=
# PROMETHEUS_SMALL_MODEL=
# Advanced: point at any compatible endpoint / bring your own key var:
# PROMETHEUS_BASE_URL=
# PROMETHEUS_API_KEY=

# ── Prometheus home (optional) ──────────────────────────────────────────────────
# Where memory (state.db), calendar.ics, outbox/ and traces/ live.
# PROMETHEUS_HOME=.prometheus

# ── Loop guardrails (optional) ──────────────────────────────────────────────
# PROMETHEUS_MAX_ITERATIONS=10
# PROMETHEUS_MAX_TOKENS=2048
# Working memory window: how many recent turns enter the prompt (default 12).
# Older turns stay in state.db + consolidation; the retrieval gate brings them
# back when relevant. Caps context/cost/latency on long (esp. Telegram) threads.
# PROMETHEUS_HISTORY_TURNS=12
# Tools this deployment does not offer, as the model names them; the prompt
# tells the model not to call them. Hosted containers set it; laptops leave it empty.
# PROMETHEUS_UNAVAILABLE_TOOLS=treg_balance,treg_resources_list

# ── Memory tuning (optional) ────────────────────────────────────────────────
# Consolidate chats into durable facts only after N new exchanges.
# PROMETHEUS_CONSOLIDATE_EVERY=6
# PROMETHEUS_RETRIEVAL_TOP_K=4

# ── GitHub, read-only (optional) ────────────────────────────────────────────
# Lets the agent read pull requests and issues through the `gh` CLI. There is
# no token here: gh keeps its own credentials, so run `gh auth login` once and
# nothing sensitive lands in this file. The tool is READ-ONLY by construction —
# it can list/view PRs and issues and read a diff, and refuses to merge,
# comment, close or push. Off by default because every registered tool ships in
# every prompt, and this is maintainer capability, not assistant capability.
# PROMETHEUS_GH_TOOL=1
# Repo to assume when a call omits one (for running outside a checkout):
# PROMETHEUS_GH_REPO=Engine-User/Prometheus-AI-Agent

# ── Open files in your editor (optional) ────────────────────────────────────
# Dashboard "reveal" links open files in this editor when it's on PATH
# (falls back to cursor, then code, then Finder):
# PROMETHEUS_EDITOR=cursor

# ── Experimental tools (optional) ───────────────────────────────────────────
# delegate_task (LIVE): hand coding jobs to pi, the open-source coding agent —
#   npm install -g --ignore-scripts @earendil-works/pi-coding-agent
# The terminal/browser/cron skeletons still just report "coming soon".
# PROMETHEUS_EXPERIMENTAL=1
# Graph workflows: triage every message through a graph (small-model classify +
# calendar peek in parallel -> quick reply, or the normal loop as a node).
# Fails open to the plain loop on any error. See prometheus/graph/.
# PROMETHEUS_GRAPH_WORKFLOWS=1
# Max seconds pi may spend on one delegated task (default 300):
# PROMETHEUS_DELEGATE_TIMEOUT=300

# ── MCP servers (optional): pip install -e '.[mcp]' ─────────────────────────
# Connect Model Context Protocol servers by creating .prometheus/mcp.json:
#   {"servers": [{"name": "fs", "command": "npx",
#                 "args": ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]}]}
# Their tools register as <server>_<tool> and appear to the agent automatically.

# ── Voice gateway (optional): pip install -e '.[voice]' ────────────────────
# Unset = push-to-talk (Enter, speak, Enter). Set a wake word — ANY phrase —
# and `make voice` becomes always-listening:
# PROMETHEUS_WAKE_WORD=prometheus prometheus
# What it says when it wakes:
# PROMETHEUS_WAKE_ACK=Yes?
# Stay in the conversation after answering (Siri-style) for N seconds before
# needing "prometheus prometheus" again:
# PROMETHEUS_FOLLOWUP_SECONDS=8
# STT model (tiny/base/small — bigger = more accurate, slower):
# PROMETHEUS_WHISPER_MODEL=base
# TTS engine: 'say' (macOS built-in, default) or 'kokoro' (neural British
# butler — first: pip install kokoro soundfile):
# PROMETHEUS_TTS=say
# PROMETHEUS_VOICE=Daniel
# (for kokoro use voices like bm_george / bm_fable)
# Mic sensitivity: lower if it never hears you, higher if it wakes on noise:
# PROMETHEUS_MIC_THRESHOLD=0.005

# ── Reliability (optional) ──────────────────────────────────────────────────
# Max seconds any single LLM call may take before erroring instead of hanging:
# PROMETHEUS_LLM_TIMEOUT=120