An AI-native browser skill that renders pages as compact, handle-addressable OwlMark text to let agents navigate, click, and type without relying on screenshots or coordinate guessing.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install owl-browser
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 owl-browser 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 Owl Browser Agent?
Owl Browser is an AI-native browser designed specifically for agentic execution. Unlike traditional browsers that require expensive visual model processing of screenshots or parsing through messy, oversized HTML, Owl Browser renders each web page as OwlMark. OwlMark is a highly compact, structured, and handle-addressable text view representing exactly what is visible on the screen. This allows AI agents to read, navigate, and act with speed and high cost-efficiency.
By leveraging Openclaw Skills, developers can integrate this agent rendering capability directly into their workflows. Rather than guessing pixel coordinates or managing complex CSS selectors, the agent interacts with simple handle tokens like l5 or x27 generated during the observation stage. This optimizes performance and drastically reduces token overhead while ensuring highly reliable automation.
Owl Browser Agent Use Cases
- Automating web research and data extraction from complex, interactive web applications.
- Integrating browser-based actions into AI agent workflows without using resource-heavy visual models.
- Performing automated form fills, button clicks, and logins using simple handle tokens.
- Navigating modern single-page applications (SPAs) and handling client-side rendering latency natively.
- Running headless web automation reliably across various Openclaw Skills pipelines.
How Owl Browser Agent Works
- Context Creation: Initialize an execution session with the
browser_create_contextendpoint, specifying therender_modeasagentto receive handle-addressable outputs. - Navigation: Command the browser to navigate to a target URL using
browser_navigate. This step loads the page but does not return content. - Observation: Execute
browser_observeto capture the page's structure. The browser returns an OwlMark text tree, a list of interactive handles, page metadata, and a token estimate. - Action Execution: Perform clicks or typing operations by target handle tokens (e.g.,
l5for a link orx27for a textbox) viabrowser_clickorbrowser_type. - Verification & Iteration: Check the returned action effect (such as
navigatedordom-changed), re-observe the screen to verify state changes, and repeat steps as necessary until the task is complete before closing the context.
Owl Browser Agent Setup
Prerequisites
Make sure you have the following environment variables configured:
OWL_API_ENDPOINT: The base URL of the Owl Browser HTTP API (e.g.,http://localhost:8080).OWL_API_TOKEN: The bearer token used to authenticate calls.- System Binaries:
curlmust be installed on your system.
Quick Integration
You can interact with the API using a helper function in your terminal shell:
owl() { curl -s -X POST "$OWL_API_ENDPOINT/execute/$1" -H "Authorization: Bearer $OWL_API_TOKEN" -H "Content-Type: application/json" -d "$2"; }
Ensure this environment is properly set up so your Openclaw Skills agent can communicate with the Owl Browser instance.
Owl Browser Agent Data Schema & Taxonomy
Output Formats & Token Schema
Owl Browser structures pages as compact OwlMark text maps with precise interactable handle taxonomies. Here is how key objects are organized:
| Element Type | OwlMark Pattern | Description |
|---|---|---|
| Links | [- link "Label" [#l5]] |
Interactive links identified by dynamic handle sequences starting with l. |
| Inputs/Buttons | [textbox "Field" [#x27 val=""]] |
Inputs, textboxes, and buttons mapped with handle tokens starting with x or b. |
| Headers/Nav | [#region:nav] |
Structured semantic landmarks for target scoping. |
Response Metadata States
When executing browser_observe, the metadata schema provides status indicators to guide the agent:
ready: Safe to interact with elements on screen.pending: Page is unhydrated or rendering dynamically. The schema provides aretry_after_msdelay value.incomplete: Chrome rendered but main content failed; requires vision fallbacks or a re-observation loop.
Owl Browser Agent Advanced Features
- Detail-Level Scoping: Dynamically scale output complexity from
min(minimalist),normal,full, tooutline(headings-only layout for massive articles). - Drill-Down Extraction: Use
browser_expandandbrowser_read_nodeto zoom in on collapsed regions or pull raw text from specific DOM nodes without bloating the primary prompt token context. - Error & Surface Recovery Hooks: Native telemetry detects when non-text elements (like
canvasorwebgl) block automation, allowing agents to selectively trigger visual fallback screenshots (render_mode: both). - Multi-Agent Profile Scoping: Integrate via Model Context Protocol (MCP) using distinct toolset profiles (
agent,automation,webdev,full) optimized for distinct Openclaw Skills tasks.
SKILL.md
Loading
Related Openclaw Skills
File Management
A robust, secure cloud file management tool for AI agents that handles uploads up to 100MB, download URLs, sharing, and meta-tagging.
Data Format Validation
A multi-domain data format validation tool for AI agents that ensures clean input processing for JSON, network configurations, financial identifiers, and general strings.
Internet Conversion OS (互联网转化操作系统)
An AI-powered full-funnel marketing operating system designed to turn product value into user action, from initial attention to continuous customer retention.
Jarvis Core
A foundational personality, memory, and cognitive architecture that transforms AI agents from transactional tools into highly empathetic, self-improving companions.
TikTok Product Launch OS (TK实拍打品OS)
An end-to-end, data-driven operating system designed to transform e-commerce products into repeatable, high-converting TikTok video campaigns.
Build Protocol Decision
A battle-tested decision-making protocol for AI agents to handle high-stakes investments, tech selections, and major purchases with rigorous real-time data verification.
Featured*
