Language: English | 简体中文

🛡️ ArbiterOS

ArbiterOS: One-Command Safety Harness for AI Agents

Autonomy under control: LLMs reason, ArbiterOS enforces.

🦾 ArbiterOS provides an ultra-lightweight, one-command installation for ArbiterOS-Kernel.

⚡ Delivers practical runtime safety and governance for agent systems while keeping local setup simple.

📊 Optional langfuse integration adds trace visualization and governance observability.

🧭 Deterministic rules for probabilistic AI, with instruction-level governance before execution.

ArbiterOS in Agent Systems

Why ArbiterOS

• Zero code intrusion for full-scope agents like OpenClaw and Nanobot.
• Instruction-flow parsing plus taint-aware data-flow policy enforcement.
• 100% support for local private deployment.
• Full lifecycle inspection, analysis, and update support for governance processes and policy strategies.

Benchmarks

ArbiterOS improves interception and warning coverage across multiple agent safety evaluations:

• Native OpenClaw (GPT + Claude): 6.17% -> 92.95%
• Agent-SafetyBench (Claude Sonnet 4): 0% -> 94.25%
• AgentDojo (GPT-4o): 0% -> 93.94%
• WildClawBench (GPT-5.2): 55% -> 100% (warning-oriented metric)

What It Does

The installer at the repository root will:

• verify required commands (curl, git) and install uv to user space
• ensure Python 3.12+ (install via uv when needed)
• clone or update ArbiterOS
• install kernel dependencies (uv sync --group dev)
• create ArbiterOS-Kernel/.env from .env.example
• guide you to fill the first model entry in ArbiterOS-Kernel/litellm_config.yaml
• update ~/.openclaw/openclaw.json for arbiteros provider and model defaults
• restart OpenClaw gateway and run openclaw dashboard
• create runnable scripts: run-kernel.sh / run-kernel.ps1

Project Structure

• ArbiterOS-Kernel: core security/governance module. Use install.sh + run-kernel.sh (or Windows equivalents) to install and run it in background.
• langfuse: optional module for Langfuse-style UI, governance details, and runtime trace visualization.

Quick Start

Get Started in 3 Steps

1. Install and start ArbiterOS-Kernel (default port: 4000).
2. Configure models, API keys, and policy rules in ArbiterOS-Kernel/litellm_config.yaml.
3. Point your agent model provider to http://127.0.0.1:4000/v1.

Install

# Linux / macOS
git clone https://github.com/cure-lab/ArbiterOS.git
cd ArbiterOS
chmod +x install.sh
./install.sh

# Windows (PowerShell)
git clone https://github.com/cure-lab/ArbiterOS.git
cd ArbiterOS
Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass
.\install-windows.ps1

Start Kernel

# Linux / macOS
./run-kernel.sh

# Windows (PowerShell)
.\run-kernel.ps1

Optional: Langfuse UI

cd ArbiterOS/langfuse
cp .env.prob.example .env
docker compose -f docker-compose.yml up -d --build

Optional: User systemd Service

If you want background auto-restart and easier operations, use a user-level service:

• service name: arbiteros-kernel
• service file: ~/.config/systemd/user/arbiteros-kernel.service
• working directory: ArbiterOS/ArbiterOS-Kernel
• start command: uv run poe litellm

Useful commands:

systemctl --user status arbiteros-kernel
journalctl --user -u arbiteros-kernel -f
systemctl --user restart arbiteros-kernel

TODO

• Support NanoBot
• Evaluate on Agent SafetyBench
• Evaluate on AgentDojo
• Evaluate on Wild Claw Bench
• Evaluate on ToolEmu
• Use skill-scanner for skill safety analysis
• Support Linux system
• Support Windows system
• Support macOS
• Protect long-term memory files in agents
• Periodically analyze consistency of role positioning, intent, and behavior
• Detect prompt injection using clustered dataflow information
• Pre-check input data
• Self-evolving policy mechanism
• Support multi-modal models