Environment Variables
GuardClaw is configured through environment variables stored in a .env file in your working directory. All variables can be set via guardclaw config set or by editing .env directly.
Quick reference
bash
# View all current settings
guardclaw config show
# Set a variable
guardclaw config set <KEY> <VALUE>LLM Backend
| Variable | Default | Description |
|---|---|---|
SAFEGUARD_BACKEND | lmstudio | Active LLM backend for risk scoring |
LMSTUDIO_URL | http://localhost:1234/v1 | LM Studio / OpenAI-compatible API URL |
LMSTUDIO_MODEL | auto | Model name (or auto to use whatever is loaded) |
LMSTUDIO_API_KEY | — | API key for LM Studio (optional) |
LLM_API_KEY | — | Generic LLM API key |
OLLAMA_URL | http://localhost:11434 | Ollama API URL |
OLLAMA_MODEL | llama3 | Ollama model name |
ANTHROPIC_API_KEY | — | Anthropic API key (for anthropic backend) |
OPENROUTER_API_KEY | — | OpenRouter API key |
OPENROUTER_MODEL | — | OpenRouter model ID (e.g. openai/gpt-4o) |
OPENAI_API_KEY | — | OpenAI API key |
GEMINI_API_KEY | — | Google Gemini API key |
KIMI_API_KEY | — | Kimi (Moonshot) API key |
MINIMAX_API_KEY | — | MiniMax API key |
Backend values
| Value | Description |
|---|---|
lmstudio | Local LLM via LM Studio (recommended) |
ollama | Local LLM via Ollama |
anthropic | Claude API |
openrouter | OpenRouter (400+ models) |
minimax | MiniMax API |
built-in | Apple Silicon MLX models |
fallback | Rule-based only, no LLM |
Cloud Judge
| Variable | Default | Description |
|---|---|---|
CLOUD_JUDGE_ENABLED | false | Enable cloud-based escalation |
CLOUD_JUDGE_MODE | local-only | Evaluation mode |
CLOUD_JUDGE_PROVIDER | — | Cloud provider for escalation |
Evaluation modes
| Mode | Description |
|---|---|
local-only | All evaluation on local LLM |
mixed | Local first, cloud escalates risky calls |
cloud-only | All evaluation via cloud API |
Cloud providers
| Provider | Description |
|---|---|
claude | Anthropic Claude (OAuth or API key) |
openai-codex | OpenAI Codex / ChatGPT (OAuth or API key) |
minimax | MiniMax (OAuth or API key) |
kimi | Kimi / Moonshot (API key) |
openrouter | OpenRouter (API key) |
gemini | Google Gemini (API key) |
openai | OpenAI (API key) |
Approval Policy
| Variable | Default | Description |
|---|---|---|
GUARDCLAW_APPROVAL_MODE | auto | How to respond to risky tool calls |
GUARDCLAW_AUTO_ALLOW_THRESHOLD | 6 | Scores at or below this are auto-allowed |
GUARDCLAW_ASK_THRESHOLD | 8 | Scores at or below this (in prompt mode) trigger user confirmation |
GUARDCLAW_AUTO_BLOCK_THRESHOLD | 9 | Scores at or above this are auto-blocked |
GUARDCLAW_MEMORY_AUTO_APPROVE | unset | Set to 1 to let explicit trusted-memory entries skip judge evaluation |
GUARDCLAW_CONSOLIDATED_MEMORY | unset | Set to 1 to load/update LLM-generated project/global memory summaries |
Approval modes
| Mode | Description |
|---|---|
auto | Score, warn agent, and flag risky calls (recommended) |
prompt | Pause execution and ask user for approval |
monitor-only | Score and log only, no intervention |
Threshold behavior
Risk scores range from 1 to 10. The three thresholds control the decision flow:
Score: 1 ──── 3 ──── 6 ──── 8 ──── 9 ──── 10
│ SAFE │ WARN │ ASK │ BLOCK │
└────────┘──────┘──────┘───────┘
auto-allow ask auto-block- Score ≤ auto-allow (default ≤ 6): automatically allowed
- Score ≤ ask (default ≤ 8): in
promptmode, asks user; inautomode, warns agent - Score ≥ auto-block (default ≥ 9): automatically blocked
Connections
| Variable | Default | Description |
|---|---|---|
BACKEND | auto | Gateway connection mode |
OPENCLAW_TOKEN | — | OpenClaw gateway authentication token |
QCLAW_TOKEN | — | Qclaw gateway authentication token |
PORT | 3002 | GuardClaw server port |
Gateway modes
| Mode | Description |
|---|---|
auto | Connect to any detected gateway |
openclaw | Connect to OpenClaw gateway only |
qclaw | Connect to Qclaw gateway only |
nanobot | Connect to nanobot gateway only |
Notifications
| Variable | Default | Description |
|---|---|---|
TELEGRAM_BOT_TOKEN | — | Telegram bot token for alert notifications |
TELEGRAM_CHAT_ID | — | Telegram chat ID to send alerts to |
DISCORD_WEBHOOK_URL | — | Discord webhook URL for alert notifications |
.env file location
The .env file is loaded from your current working directory when starting GuardClaw. This allows per-project configuration.
bash
# In project A
cd ~/projects/project-a
guardclaw start
# Reads ~/projects/project-a/.env
# In project B
cd ~/projects/project-b
guardclaw start
# Reads ~/projects/project-b/.envExample .env
ini
# LLM Backend
SAFEGUARD_BACKEND=lmstudio
LMSTUDIO_URL=http://localhost:1234/v1
LMSTUDIO_MODEL=auto
# Cloud Judge (mixed mode)
CLOUD_JUDGE_ENABLED=true
CLOUD_JUDGE_MODE=mixed
CLOUD_JUDGE_PROVIDER=claude
# Approval Policy
GUARDCLAW_APPROVAL_MODE=auto
GUARDCLAW_AUTO_ALLOW_THRESHOLD=6
GUARDCLAW_ASK_THRESHOLD=8
GUARDCLAW_AUTO_BLOCK_THRESHOLD=9
# Server
PORT=3002
# Agent Connections
OPENCLAW_TOKEN=eyJ0eXAiOiJKV1Qi...