功能描述

Full-featured headless browser for OpenClaw agents. Navigate, snapshot with accessibility tree (@ref clicks), tabs, JS execution, cookie import. No vision mo...

使用说明 (SKILL.md)

Gbrow — The Browser Your AI Agent Actually Needs

Name: Gbrow
Author: ashish797

A full-featured headless browser powered by Playwright and Bun. Uses the accessibility tree for page reading — not expensive vision models.

Why Gbrow?

Traditional (screenshots + vision)	Gbrow (accessibility tree)
Screenshot → upload to GPT-4o → wait → read	`ariaSnapshot()` → instant structured text
~$0.01 per page read	Free
3-10 seconds per page	\x3C 100ms
Fails on API key issues	Always works
Click by fragile CSS selector	Click by @ref (`@e1`, `@e2`, etc.)

Quick Setup

# Clone and install
git clone https://github.com/ashish797/Gbrow.git ~/.openclaw/workspace/skills/Gbrow
cd ~/.openclaw/workspace/skills/Gbrow
bash setup.sh

Or one-liner:

curl -fsSL https://raw.githubusercontent.com/ashish797/Gbrow/main/setup.sh | bash

How It Works

1. Start the server

cd ~/.openclaw/workspace/skills/Gbrow
bun run src/server.ts

2. Read the page (accessibility tree)

The snapshot gives you a structured view with clickable refs:

@e1 [heading] "Welcome" [level=1]
@e2 [link] "Get Started"
@e3 [button] "Sign in"
@e4 [textbox] "Search"

3. Click by ref

click @e2     → clicks "Get Started"
fill @e4 "query"  → types into search box

Commands

Navigation

Command	Description	Example
`goto \x3Curl>`	Navigate to URL	`goto https://example.com`
`back`	History back	`back`
`forward`	History forward	`forward`
`reload`	Reload page	`reload`
`url`	Print current URL	`url`

Reading

Command	Description	Example
`snapshot`	Accessibility tree with @refs	`snapshot -i` (interactive only)
`text`	Cleaned page text	`text`
`html [selector]`	Raw HTML	`html .article`
`links`	All links as "text → href"	`links`
`forms`	Form fields as JSON	`forms`

Interaction

Command	Description	Example
`click \x3Cref>`	Click element	`click @e3`
`fill \x3Cref> \x3Ctext>`	Fill input	`fill @e4 "hello"`
`select \x3Cref> \x3Cvalue>`	Select dropdown	`select @e5 "option1"`
`type \x3Cref> \x3Ctext>`	Type with keyboard	`type @e4 "search term"`
`press \x3Ckey>`	Press key	`press Enter`
`scroll \x3Cdirection>`	Scroll page	`scroll down`

Inspection

Command	Description	Example
`js \x3Cexpr>`	Run JavaScript	`js document.title`
`css \x3Csel> \x3Cprop>`	Computed CSS	`css .box color`
`attrs \x3Cref>`	Element attributes	`attrs @e1`
`is \x3Cprop> \x3Cref>`	State check	`is visible @e3`

Tabs

Command	Description
`tabs`	List open tabs
`tab N`	Switch to tab N
`newtab`	Open new tab
`closetab`	Close current tab

Visual

Command	Description
`screenshot`	Take screenshot
`responsive \x3Cw> \x3Ch>`	Set viewport size
`pdf`	Save page as PDF

Snapshot Flags

Flag	Description
`-i`	Interactive elements only (buttons, links, inputs)
`-c`	Compact (remove empty structural nodes)
`-d N`	Limit tree depth
`-s \x3Csel>`	Scope to CSS selector
`-D`	Diff against previous snapshot
`-a`	Annotated screenshot with ref overlays

HTTP API

All commands go through the HTTP API:

# Get port and token from state file
PORT=$(python3 -c "import json; print(json.load(open('.gstack/browse.json'))['port'])")
TOKEN=$(python3 -c "import json; print(json.load(open('.gstack/browse.json'))['token'])")

# Send command
curl -s -X POST "http://127.0.0.1:${PORT}/command" \
  -H "Authorization: Bearer ${TOKEN}" \
  -H "Content-Type: application/json" \
  -d '{"command":"goto","args":["https://example.com"]}'

Architecture

┌─────────────┐     HTTP      ┌──────────────────┐
│  OpenClaw   │ ──────────▶  │  Gbrow Server    │
│  Agent      │              │  (Bun + Playwright)│
└─────────────┘              └────────┬─────────┘
                                      │
                                      ▼
                              ┌──────────────────┐
                              │  Chromium         │
                              │  (headless)       │
                              └──────────────────┘
                                      │
                                      ▼
                              ┌──────────────────┐
                              │ Accessibility     │
                              │ Tree (ariaSnapshot)│
                              └──────────────────┘

No vision models. No API calls. Just structured text from the browser's accessibility layer.

Credits

Built on top of gstack by Gary Tan (Y Combinator). Adapted for OpenClaw with permission under MIT license.

License

MIT

安全使用建议

What to consider before installing: - Review the install script instead of piping it: the SKILL.md recommends 'curl | bash' for setup. Copy the setup.sh URL and inspect its contents before running, or clone the repository and run scripts manually. - Run in an isolated environment: because the skill installs Bun/Chromium and runs a persistent server that stores tokens and browser profiles under ~/.gstack, prefer a throwaway VM or container so it cannot access your real browser profile or other files. - Inspect cookie-import and extension-write behavior: the code includes cookie-import logic and will attempt to locate extension directories and may write .auth.json there. If you do not want the skill reading local browser cookies or modifying extension directories, do not install it or audit/disable those code paths first. - Check exposed endpoints and logs: the server keeps large console/network buffers and exposes HTTP endpoints for commands and activity streaming. Confirm authentication and that logs do not leak sensitive data before allowing other agents/processes access. - Principle of least privilege: avoid running the install as root and avoid giving the server access to sensitive filepaths. If you must run it locally, consider running it under a restricted user and firewalling the server port to localhost only. - If unsure, ask the maintainer for a reproducible package or an audited release: prefer installing from a pinned, reviewed release (GitHub release archive) instead of a raw branch tip, and request clear docs on cookie import and extension writes. Why 'suspicious' not 'malicious': the code is consistent with a Playwright-based browser skill and many behaviors are explainable, but there are multiple mismatches and higher-privilege file operations (cookie import, writing .auth.json to discovered extension directories, piping remote install scripts) that are significant enough to recommend caution and manual review before trusting the skill.

功能分析

Type: OpenClaw Skill Name: gbrow Version: 1.0.0 The Gbrow skill bundle provides a headless browser with high-risk capabilities, most notably a module (src/cookie-import-browser.ts) that contains logic to locate and decrypt local browser cookies (Chrome, Edge, Brave, etc.) on macOS and Linux using system keychain utilities. While framed as a 'Cookie Picker' feature for session persistence, this functionality is a common characteristic of info-stealer malware. Additionally, the bundle includes tools for arbitrary JavaScript execution (src/read-commands.ts) and a background agent (src/sidebar-agent.ts) that spawns the 'claude' CLI with Bash and filesystem access. Although the code implements some security controls like URL validation (src/url-validation.ts) and log redaction (src/activity.ts), the combination of credential decryption and shell execution represents a significant security risk.

能力评估

⚠ Purpose & Capability

The name/description (headless browser using Playwright) matches the code. However the registry entry declares no required env vars or config paths while the implementation reads/writes state under .gstack, uses environment variables (BROWSE_STATE_FILE, BROWSE_EXTENSIONS_DIR, CI/CONTAINER, etc.), and includes a cookie-import component and extension-write behavior that are higher privilege than the registry metadata indicates.

⚠ Instruction Scope

SKILL.md instructs cloning the GitHub repo and piping setup.sh via curl | bash and to run a persistent local server. The instructions reference reading the local state file (.gstack/browse.json) to obtain a token and port. The docs omit some high-impact behaviors present in code: writing .auth.json into found extension directories, cookie import logic, and large network/console buffers and activity streaming endpoints that may contain sensitive data. Pre-scan found a 'system-prompt-override' pattern in SKILL.md (prompt-injection signal).

⚠ Install Mechanism

There is no registry install spec, but SKILL.md/setup.sh instructs executing remote scripts: curl https://raw.githubusercontent.com/…/setup.sh | bash and curl https://bun.sh/install | bash, plus npx playwright install chromium. These downloads are from well-known hosts (GitHub, bun.sh, playwright via npx) but piping remote shells to bash and downloading browser binaries are moderate-to-high risk operations and worth manual review before execution.

⚠ Credentials

The skill declares no required credentials or config paths but the code uses multiple env vars (BROWSE_EXTENSIONS_DIR, BROWSE_STATE_FILE, BROWSE_SERVER_SCRIPT, HOME, CI/CONTAINER) and writes state/tokens into ~/.gstack. It also contains cookie-import functionality (likely reads browser cookie DBs) and will write .auth.json into extension directories it finds, which can expose tokens to other components. The undeclared access to user browser cookies and potential writes into other skills' extension directories are disproportionate and sensitive.

⚠ Persistence & Privilege

The server launches a persistent local daemon, creates ~/.gstack state and a persistent Chromium profile, and can run detached in background. Importantly, in headed mode it will locate an extension path and may write a .auth.json file there (modifying a discovered extension directory). That behavior modifies other on-disk components and can expose tokens — this matches the rule 'modify OTHER skills' configurations' and is a notable risk. The skill does not set always:true, but it does request persistent local presence through files and background processes.

版本历史

v1.0.0

Initial release: Full-featured headless browser for OpenClaw agents. Accessibility tree snapshot, @ref clicks, tabs, JS execution. No vision model needed.

元数据

Slug gbrow

版本 1.0.0

许可证 MIT-0

累计安装 0

当前安装数 0

历史版本数 1

常见问题

Gbrow 是什么？

Full-featured headless browser for OpenClaw agents. Navigate, snapshot with accessibility tree (@ref clicks), tabs, JS execution, cookie import. No vision mo... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 103 次。

如何安装 Gbrow？

在 OpenClaw 或 Claude Code 对话框中运行命令「/install gbrow」即可一键安装，无需额外配置。

Gbrow 是免费的吗？

是的，Gbrow 完全免费，采用 MIT-0 许可证，可自由下载、安装和使用。

Gbrow 支持哪些平台？

Gbrow 跨平台运行，可在任意部署了 OpenClaw / Claude Code 的环境中使用（cross-platform）。

谁开发了 Gbrow？

由 ashish797（@ashish797）开发并维护，当前版本 v1.0.0。

Gbrow