A decision-brokering skill that allows AI agents to push complex plans and architectural choices for asynchronous human review and approval.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install arbiter
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install arbiter using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Arbiter?
The Arbiter skill provides a robust bridge between autonomous AI agents and human oversight within the Openclaw Skills ecosystem. It is designed for scenarios where an agent requires human judgment on high-stakes decisions, such as selecting a software architecture, approving a multi-step implementation plan, or choosing between technical tradeoffs. By utilizing this skill, agents can maintain progress without needing real-time synchronous interaction, queuing requests for a human to review at their convenience.
Built to enhance the reliability of agentic workflows, Arbiter ensures that Openclaw Skills can operate with a safety net. It allows developers to define clear decision points where human intervention is mandatory, thereby reducing the risk of autonomous errors in production environments. The skill manages the full lifecycle of a decision, from initial push to final notification, making it an essential tool for sophisticated agent management.
Arbiter Use Cases
- Reviewing complex implementation plans before an agent begins writing code.
- Making architectural decisions that involve significant technical tradeoffs or infrastructure costs.
- Requesting human judgment when an agent encounters ambiguous requirements that cannot be resolved through research.
- Batching multiple related decisions together to minimize human interruption and context switching.
How Arbiter Works
- The AI agent identifies a decision point and uses the arbiter_push tool to submit a structured JSON decision plan.
- The request is saved as a Markdown file with YAML frontmatter in a local queue, making it accessible to the Arbiter Zebu review interface.
- A human reviewer examines the plan, selects the preferred options or provides custom feedback asynchronously.
- The agent can periodically check the status of the plan using arbiter_status or use arbiter_await to block until the human responds.
- Once the human completes the review, the agent retrieves the answers via arbiter_get and proceeds with the approved course of action.
Arbiter Setup
Install the skill using one of the following methods to integrate it with your Openclaw Skills setup:
# Quick install via ClawHub
clawhub install arbiter
# Or via bun for global CLI access
bun add -g arbiter-skill
# Manual installation from source
git clone https://github.com/5hanth/arbiter-skill.git
cd arbiter-skill && npm install && npm run build
ln -s $(pwd) ~/.clawdbot/skills/arbiter
Ensure you have the Arbiter Zebu bot running and the ~/.arbiter/queue/ directory created.
Arbiter Data Schema & Taxonomy
The Arbiter skill organizes data within the ~/.arbiter/ directory using the following structure:
| Path | Purpose |
|---|---|
~/.arbiter/queue/pending/ |
Contains plans awaiting human review in Markdown/YAML format. |
~/.arbiter/queue/completed/ |
Stores archived plans after they have been answered. |
~/.arbiter/queue/notify/ |
Holds session-specific notification files for agents to process completion. |
Decision plans are defined by a title, priority, and an array of decision objects, each containing a unique ID and a list of options for the reviewer.
Arbiter Advanced Features
- Automatic agent and session detection via CLAWDBOT_AGENT and CLAWDBOT_SESSION environment variables.
- Priority-based queuing support (low, normal, high, urgent) to help human reviewers triage requests.
- Flexible decision types including multiple-choice options and free-text custom answers.
- Notification hooks that allow blocked agents to be woken up once a human has provided input.
- Support for batching related decisions into a single plan to provide better context for the reviewer.
SKILL.md
Loading
Related Openclaw Skills
Public API Skill Creator
A specialized tool to discover, test, and transform public APIs into ready-to-use Openclaw Skills.
Yahoo Finance Data (yahooquery)
A comprehensive Python library providing programmatic access to Yahoo Finance endpoints including real-time pricing, fundamentals, and analyst estimates.
Adaptive Skill Factory
A procedural skill for agents to dynamically manage their capabilities by discovering, installing, or building targeted skills.
OpenClaw Marketplace (openclawmp)
OpenClaw Marketplace is an asset registry and ecosystem where AI agents and developers discover, install, and share specialized capabilities.
Singapore Property Scraper
An advanced property search tool that scrapes Singapore rental and sale listings with support for MRT proximity and commute time filtering.
Find Profitable Stocks
A financial analysis tool that identifies high-quality companies by calculating Free Cash Flow and aggregate health scores.
METADATA
Featured*
