Configuration

The guardclaw config command manages all GuardClaw settings. Run it without arguments for an interactive menu, or use subcommands for direct access.

guardclaw config [subcommand]

Subcommands

Subcommand	Description
(none) / menu	Interactive configuration menu
setup	Re-run the setup wizard
eval	Change evaluation mode
llm	Change LLM backend
mode	Change approval mode
thresholds	Change risk thresholds
agents	Manage agent connections
set <KEY> <VALUE>	Set any environment variable
show	Show all current settings
set-token <token>	Set OpenClaw gateway token
get-token	Show OpenClaw gateway token
detect-token	Auto-detect OpenClaw token

All changes are hot-reloaded to the running server — no restart required.

Interactive menu

guardclaw config

Opens a menu with all configuration categories. Use arrow keys to navigate, Enter to select.

Setup wizard

guardclaw setup
# or
guardclaw config setup

Re-runs the 3-step first-run wizard:

1. Cloud judge — provider, OAuth/API key, and model
2. Response mode — how to respond to risky tool calls
3. Agent connections — detect and connect AI agents

Evaluation mode

guardclaw config eval

Choose how risk evaluation is performed:

Mode	Description
local-only	All evaluation done by local LLM — private, fast
mixed	Local LLM first, cloud escalates risky calls (recommended)
cloud-only	All evaluation via cloud API

In mixed mode, tool calls that score above a configurable threshold are re-evaluated by a cloud LLM for a second opinion. This combines the speed of local evaluation with the accuracy of larger cloud models.

LLM backend

guardclaw config llm

Interactive picker for the LLM backend used for local risk scoring. Shows the current backend and lets you switch.

Available backends

Backend	Value	Description
LM Studio	lmstudio	Local LLM via LM Studio (default, recommended)
MiniMax	minimax	MiniMax API
Ollama	ollama	Local LLM via Ollama
Anthropic	anthropic	Claude API
OpenRouter	openrouter	400+ models via OpenRouter
Built-in	built-in	Apple Silicon MLX models (downloaded & managed by GuardClaw)
Fallback	fallback	Rule-based only, no LLM calls

LM Studio configuration

When selecting lmstudio, you'll be prompted for:

• URL — LM Studio API endpoint (default: http://localhost:1234/v1)
• API Key — optional, if your LM Studio requires authentication
• Model — choose from detected models or enter manually (default: auto)

The CLI auto-fetches available models from your running LM Studio instance.

Recommended model

qwen/qwen3-4b-2507 — fast, accurate, and small enough to run alongside your AI agent.

Ollama configuration

When selecting ollama:

• URL — Ollama API endpoint (default: http://localhost:11434)
• Model — choose from installed models or enter manually (default: llama3)

Anthropic configuration

When selecting anthropic:

• API Key — your Anthropic API key

OpenRouter configuration

When selecting openrouter:

• API Key — your OpenRouter API key
• Model — choose from a curated list of popular models

Available OpenRouter models include:

• openai/gpt-4o, openai/gpt-4o-mini
• anthropic/claude-sonnet-4-6, anthropic/claude-haiku-4-5
• google/gemini-2.0-flash-001
• meta-llama/llama-3.3-70b-instruct
• deepseek/deepseek-r1
• qwen/qwen3-235b-a22b

MiniMax configuration

When selecting minimax:

• API Key — your MiniMax API key (or OAuth login)
• Model — choose from available MiniMax models

Available models: MiniMax-M2.7, MiniMax-M2.7-highspeed, MiniMax-M2.5, MiniMax-M2.5-highspeed, MiniMax-M2.1, MiniMax-M1, MiniMax-Text-01.

Built-in (MLX)

The built-in backend runs Apple Silicon MLX models directly. Models are downloaded and managed via the dashboard or CLI after the server starts.

guardclaw model list           # List available models
guardclaw model load <id>      # Load a model
guardclaw model unload         # Unload the current model

See Built-in Backend for details.

Fallback

Rule-based scoring only — no LLM calls. Uses pattern matching and heuristics. Suitable when no LLM is available or for testing.

Direct backend switching

guardclaw config set SAFEGUARD_BACKEND ollama

Approval mode

guardclaw config mode

Choose how GuardClaw responds to risky tool calls:

Mode	Description
auto	Score, warn, and flag risky calls to the agent (recommended)
prompt	Pause and ask the user for approval on risky calls
monitor-only	Score and log only, no intervention

Direct mode switching

guardclaw config set GUARDCLAW_APPROVAL_MODE auto

Thresholds

guardclaw config thresholds

Configure the three risk score thresholds that control GuardClaw's behavior:

Threshold	Default	Description
Auto-allow	6	Scores at or below this are auto-allowed
Ask	8	Scores at or below this (in prompt mode) trigger user confirmation
Auto-block	9	Scores at or above this are auto-blocked

Risk scores range from 1 to 10:

• 1–3: SAFE (green) — routine operations
• 4–7: WARNING (yellow) — review recommended
• 8–10: HIGH RISK (red) — dangerous operations

Direct threshold changes

guardclaw config set GUARDCLAW_AUTO_ALLOW_THRESHOLD 5
guardclaw config set GUARDCLAW_ASK_THRESHOLD 7
guardclaw config set GUARDCLAW_AUTO_BLOCK_THRESHOLD 9

Agent connections

guardclaw config agents

Interactive agent connection manager. Auto-detects AI agents installed on your system and lets you connect them.

Detected agents

Agent	Type	Detection
Claude Code	Hook	~/.claude directory exists
Codex	Hook	~/.codex directory exists
Gemini CLI	Hook	~/.gemini directory exists
Cursor	Hook	~/.cursor directory exists
Copilot CLI	Hook	~/.copilot directory exists
OpenCode	Hook	Binary in PATH
OpenClaw	WebSocket	~/.openclaw/openclaw.json with token
Qclaw	WebSocket	~/.qclaw/openclaw.json with token

Hook-based agents are connected by installing hook scripts into the agent's settings. See Hooks.

WebSocket agents are connected by saving their gateway token. See Tokens.

Set any variable

guardclaw config set <KEY> <VALUE>

Set any environment variable in the .env file. The change is hot-reloaded to the running server when applicable.

Hot-reloaded keys

The following key groups trigger automatic hot-reload:

Key group	Hot-reload endpoint
SAFEGUARD_BACKEND, LMSTUDIO_URL, LMSTUDIO_MODEL, LMSTUDIO_API_KEY, LLM_API_KEY, OLLAMA_URL, OLLAMA_MODEL, OPENROUTER_API_KEY, OPENROUTER_MODEL	LLM config
OPENCLAW_TOKEN	OpenClaw connection
QCLAW_TOKEN	Qclaw connection
GUARDCLAW_APPROVAL_MODE	Approval mode
GUARDCLAW_AUTO_ALLOW_THRESHOLD, GUARDCLAW_ASK_THRESHOLD, GUARDCLAW_AUTO_BLOCK_THRESHOLD	Thresholds

Keys not in the above groups are saved to .env but require a restart to take effect.

Examples

# Switch LLM backend
guardclaw config set SAFEGUARD_BACKEND openrouter

# Set OpenRouter API key
guardclaw config set OPENROUTER_API_KEY sk-or-...

# Change approval mode
guardclaw config set GUARDCLAW_APPROVAL_MODE prompt

# Set server port
guardclaw config set PORT 4000

# Set Telegram notifications
guardclaw config set TELEGRAM_BOT_TOKEN 123456:ABC...
guardclaw config set TELEGRAM_CHAT_ID -100123456789

Show settings

guardclaw config show

Display all current configuration values organized into sections:

LLM Backend

Variable	Description
SAFEGUARD_BACKEND	Active backend
LMSTUDIO_URL	LM Studio API URL
LMSTUDIO_MODEL	LM Studio model name
LMSTUDIO_API_KEY	LM Studio API key
OLLAMA_URL	Ollama API URL
OLLAMA_MODEL	Ollama model name
ANTHROPIC_API_KEY	Anthropic API key
OPENROUTER_API_KEY	OpenRouter API key
OPENROUTER_MODEL	OpenRouter model
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

Cloud Judge

Variable	Description
CLOUD_JUDGE_ENABLED	Cloud judge on/off
CLOUD_JUDGE_MODE	Mode: local-only, mixed, cloud-only
CLOUD_JUDGE_PROVIDER	Provider: claude, openai-codex, minimax, kimi, openrouter, gemini, openai

Approval Policy

Variable	Description
GUARDCLAW_APPROVAL_MODE	Mode: auto, prompt, monitor-only
GUARDCLAW_AUTO_ALLOW_THRESHOLD	Auto-allow threshold (scores ≤ this)
GUARDCLAW_ASK_THRESHOLD	Ask threshold (scores ≤ this in prompt mode)
GUARDCLAW_AUTO_BLOCK_THRESHOLD	Auto-block threshold (scores ≥ this)

Connections

Variable	Description
BACKEND	Gateway mode: auto, openclaw, qclaw, nanobot
OPENCLAW_TOKEN	OpenClaw gateway token
QCLAW_TOKEN	Qclaw gateway token
PORT	Server port (default: 3002)

Notifications

Variable	Description
TELEGRAM_BOT_TOKEN	Telegram bot token
TELEGRAM_CHAT_ID	Telegram chat ID
DISCORD_WEBHOOK_URL	Discord webhook URL

Agent Hooks

Also shows the connection status of all detected agents (connected/not connected).

Tokens

Set token

guardclaw config set-token <token>

Set the OpenClaw gateway token. Equivalent to guardclaw config set OPENCLAW_TOKEN <token>, but with a dedicated command for convenience. Hot-reloaded to the running server.

Get token

guardclaw config get-token

Display the current OpenClaw token (masked for security):

🔑 OPENCLAW_TOKEN: eyJ0eXAi...aGVk
   Full: eyJ0eXAiOi...

If no token is set:

ℹ️  No OPENCLAW_TOKEN set.
   Set: guardclaw config set-token <token>
   Auto-detect: guardclaw config detect-token

Detect token

guardclaw config detect-token [--save]

Auto-detect the OpenClaw token from ~/.openclaw/openclaw.json.

Flag	Description
--save, -s	Save the detected token to .env and hot-reload

Without --save, it only prints the found token and shows how to save it:

✅ Found: eyJ0eXAiOiJKV1Qi...
   To save: guardclaw config set-token eyJ...
   Or:      guardclaw config detect-token --save