Tinman is a security-focused research agent that protects AI systems from failure modes through proactive scanning, real-time tool-call validation, and synthetic attack probing.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install agent-tinman
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 agent-tinman 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 Tinman?
Tinman serves as a forward-deployed research and protection layer for AI agents. By integrating this into Openclaw Skills, developers can discover unknown failure modes through systematic experimentation and automated security checks. It acts as a safety barrier by analyzing tool calls before execution, classifying risks by severity, and proposing mitigations mapped to internal controls like SOUL.md.
This skill is designed for developers who need to ensure their autonomous agents operate within safe boundaries. It provides a robust framework for detecting prompt injection, tool misuse, and context bleed while maintaining local privacy and performance. By utilizing Tinman within your ecosystem of Openclaw Skills, you create a self-policing environment that hardens your AI against sophisticated attacks.
Tinman Use Cases
- Preventing prompt injection and instruction overrides in real-time.
- Auditing recent session histories for sensitive data leaks or context bleed.
- Running proactive security sweeps with synthetic attack probes to test agent resilience.
- Automating agent self-policing by validating tool arguments against security policies.
- Monitoring agent behavior through local event streaming and specialized security dashboards.
How Tinman Works
- Initializes a local workspace to manage security patterns, allowlists, and configuration files.
- Intercepts or checks tool calls via the check command to provide immediate security verdicts: SAFE, REVIEW, or BLOCKED.
- Scans session history periodically to identify post-execution failure modes and categorize them by severity levels from S0 to S4.
- Executes proactive security sweeps using hundreds of synthetic probes to identify vulnerabilities in the agent logic or sandbox configuration.
- Streams structured security events to local JSONL files for real-time visualization and third-party dashboard integration.
Tinman Setup
To get started with this security layer for Openclaw Skills, install the required packages and initialize your workspace:
pip install AgentTinman>=0.2.1 tinman-openclaw-eval>=0.3.2
/tinman init
To enable autonomous protection, update your agent SOUL.md to require a security check before executing sensitive tools like bash, read, or write.
Tinman Data Schema & Taxonomy
Tinman organizes security data within the local workspace directory (~/.openclaw/workspace/):
| File | Purpose |
|---|---|
tinman.yaml |
Main configuration for security modes, focus areas, and severity thresholds. |
tinman-events.jsonl |
Local event stream for real-time monitoring and dashboard bridges. |
tinman-findings.md |
Actionable report containing session analysis and detected failure modes. |
tinman-sweep.md |
Detailed results from synthetic attack probe execution. |
tinman-findings.md |
Comprehensive list of identified risks and proposed mitigations. |
Tinman Advanced Features
- Multi-mode security profiles (Safer, Risky, YOLO) to balance agent autonomy with strict protection.
- Continuous monitoring via WebSocket gateway integration for instant event alerts.
- Extensive library of 288 synthetic attack probes covering exfiltration, privilege escalation, and supply chain attacks.
- Local Oilcan bridge integration for human-readable status dashboards and plain-language setup.
- Support for scheduled security auditing via heartbeat job integration.
SKILL.md
Loading
Related Openclaw Skills
Linkfuse
A specialized skill for generating affiliate short links through the Linkfuse REST API, designed for seamless AI-driven marketing automation.
Meeting Notes Pro
A comprehensive meeting productivity toolkit designed to transform raw notes into structured outcomes and eliminate unnecessary meetings.
Agile Toolkit
An AI-powered Agile Coach that helps teams facilitate ceremonies, write high-quality user stories, and optimize sprint performance through data-driven metrics.
CoinGecko Crypto Market Data
Access comprehensive cryptocurrency market data, including prices, trends, and token contract lookups, directly within your AI agent workflow.
War Intelligence Monitor
A real-time conflict monitoring and emergency alert system providing location-based threat assessments and evacuation logistics.
Alter Action Trigger
A powerful bridge for triggering over 84 specialized Alter macOS AI actions directly from your terminal or AI agent.
METADATA
Featured*
