.env.example

# HermesClaw environment variables.
# Copy to .env and fill in:
#   cp .env.example .env
#
# .env is gitignored — your secrets stay on your machine.

# ── Model ─────────────────────────────────────────────────────────────────────
# Filename inside models/ to load
MODEL_FILE=your-model-name.gguf

# ── GPU / compute ─────────────────────────────────────────────────────────────
# Number of model layers to offload to GPU.
#   0  = CPU only (works anywhere, no NVIDIA required)
#   99 = all layers on GPU (fast, requires VRAM)
#   Mac Metal: use 99 — llama.cpp handles Metal automatically
N_GPU_LAYERS=0

# Context window size in tokens.
# IMPORTANT: Must be >= 32768. Hermes system prompt + skills list is ~11k tokens.
# Values below 32768 cause "context length exceeded" on every query.
CTX_SIZE=32768

# ── Ports ─────────────────────────────────────────────────────────────────────
# llama.cpp listen port (default: 8080)
LLAMA_PORT=8080

# HermesClaw gateway webhook port (default: 8090)
HERMESCLAW_PORT=8090

# ── Messaging gateway tokens ───────────────────────────────────────────────────
# Obtain from:
#   Telegram: https://t.me/BotFather
#   Discord:  https://discord.com/developers/applications
#   Slack:    https://api.slack.com/apps
#
# Leave blank to disable that platform's gateway.
TELEGRAM_BOT_TOKEN=
DISCORD_BOT_TOKEN=
SLACK_BOT_TOKEN=

# ── Privacy and safety ────────────────────────────────────────────────────────
# Sensitivity-based inference routing threshold.
#   0.0 = always route to local model (most private, default)
#   0.7 = route low-sensitivity queries to cloud, high-sensitivity to local
#   1.0 = always route to cloud model (least private)
HERMES_PRIVACY_THRESHOLD=0.0

# Dangerous command approval mode.
#   manual — always prompt for approval (safest for unattended gateway mode)
#   smart  — uses LLM to assess risk, auto-approves safe commands (default)
#   off    — disable approval checks entirely (not recommended)
HERMES_APPROVAL_MODE=smart