OpenCLI Browser Driving Skill
Let AI agents drive your logged-in Chrome via CLI, navigating pages, filling forms, clicking, and extracting data with structured, reliable commands.
The skill instructs the agent to control a real browser via CDP, allowing read/write of page data and simulating logged-in states. The documentation clearly requires the agent to use structured commands rather than eval for write operations, and bind does not own user windows or close user tabs, showing respect for privileges. However, there is no explicit handling of sensitive data (e.g., passwords, cookies) or rollback/recovery mechanisms. Security relies on local daemon and Chrome extension, with installation source dependence. Publisher unverified, but documentation and project maintenance habits may reinforce trust. Therefore trust 10/25: main risks visible but isolation and rollback incomplete.
SKILL.md is self-consistent and detailed, including error codes and match_level feedback to guide the agent. However, these are static text descriptions; actual CLI behavior cannot be confirmed. Commands/examples may be outdated or unsynced. Static review did not execute any commands, so key-path reproduction is unverified. No dedicated tests for this skill path. Reliability 8/10 (static cap): happy path plausible but edge cases and failure feedback evidence thin.
Skill fits scenarios requiring real browser driving, such as checking notifications, filling forms, extracting data. Documentation clearly delineates boundary with opencli-adapter-author, and provides numeric refs + CSS selector contract for targeting. Environment: project supports Chinese docs and many Chinese site adapters (Bilibili, Zhihu, Xiaohongshu), indicating Chinese user friendliness. Core function depends on local Chrome and extension, not overseas services, so reachability from mainland-China is good. However, specific trigger conditions are semantic and untested. Adaptability 11/15: scenario clear but evidence for boundaries and trigger conditions limited.
SKILL.md has clear IA with prerequisites, session lifecycle, target contract, error handling, cost guide, and examples for progressive disclosure. README provides install instructions and skill directory, with Apache-2.0 license. However, dedicated skill doc lacks explicit version/changelog; maintenance responsibility only implied in README. Known limitations not fully listed (e.g., cross-origin iframe best-effort but others not). Convention 10/15: readable and usable but with hidden assumptions and governance gaps.
Static review cannot verify actual effectiveness. Documentation claims various tasks and provides examples, but no verifiable third-party execution evidence. Although e2e tests exist, they target adapters and article download, not the core browser-driving path of this skill. Effectiveness 5/7 (static cap): completes core task likely but evidence for output usability and comparative benefit limited.
Static review only sees docs and test files, cannot execute commands. CI workflows exist with e2e tests, but those cover other skill paths (article download, auth failure), not directly this skill's core driving functionality. Verifiability 3/5: auditable primary material but thin coverage.
- This skill can drive a real browser and simulate logged-in states, potentially accessing sensitive data (e.g., personal account info). Ensure explicit user consent and least privilege before execution.
- Documentation lacks explicit handling of sensitive data (passwords, cookies) and rollback mechanisms. Consider confirmation before sensitive operations.
- Skill depends on local Chrome and extension; ensure trusted installation source (official store or GitHub Releases).
- Static assessment cannot verify actual behavior; test in isolated environment before real use.
What does this skill do, and when should you use it?
The OpenCLI browser skill equips AI agents with the ability to drive a real Chrome window through a command-line interface. Agents can navigate to URLs, inspect page structure, fill forms, click elements, extract data, and intercept network requests. The skill emphasizes a selector-first target contract, structured envelopes with match-level confidence, and error codes for autonomous, reliable operation. It supports binding existing tabs, background windows, cross-origin iframes, compound form controls (dates, selects, files), and network traffic inspection.
The skill drives a live Chrome session via opencli browser commands: opens URLs, fetches state snapshots with numeric refs, finds elements by CSS or semantic locators, clicks, types, fills, selects, uploads files, drags, presses keys, scrolls, waits for conditions, extracts Markdown content, evaluates read-only JavaScript (including cross-origin iframes), captures and inspects network requests, and manages tabs and sessions. All commands return structured JSON envelopes with match counts, confidence levels, and machine-readable error codes.
- When an AI agent needs to automate form filling and submission on a site where you are logged in, such as replying to emails or posting on social platforms.
- When you need to scrape data from pages requiring authentication or interaction (e.g., personal timelines, private notifications, paginated lists), preferring to intercept their internal APIs rather than scraping the DOM.
- When you want an AI agent to drive your already-open tabs and perform multi-step flows without disrupting your login state, including parallel tasks across tabs.
- When debugging website automation: comparing DOM vs. accessibility-tree snapshots or troubleshooting element selectors that fail.
- When you need an AI agent to extract content from long articles in chunks or interact with custom React dropdowns (e.g., date pickers, category selectors).
What are this skill's strengths and limitations?
- Uses your logged-in browser, avoiding cookie copying and repeated logins.
- Structured JSON envelopes with confidence and actionable error codes, ideal for autonomous agents.
- Resilient to DOM changes via numeric refs and fingerprinting.
- Network interception preferred over DOM scraping, more reliable and cost-effective.
- Supports binding existing tabs, background windows, and cross-origin iframes.
- Requires installing multiple components: OpenCLI runtime, Browser Bridge extension, and the skill itself.
- Depends on Chrome's debugging port, vulnerable to interfering extensions like 1Password.
- Automation can be brittle; some sites require custom adapters beyond this skill.
- No built-in test suite; reliability relies on `opencli doctor` diagnostics.
- `eval` is read-only and cannot submit forms; must use structured commands.
How do you install this skill?
- Install OpenCLI: recommended via the desktop app (opencli.info/download), or globally via npm with
npm install -g @jackwener/opencli(requires Node.js >= 20). - Install the Browser Bridge extension: from the Chrome Web Store, or download the zip from GitHub Releases and load it in chrome://extensions.
- Add the skill to your AI agent: run
npx skills add jackwener/opencli --skill opencli-browser. - Verify with
opencli doctor; ensure it is green.
How do you use this skill?
Ensure Chrome is running and opencli doctor passes. Describe a task to your AI agent, e.g., "Help me check my Xiaohongshu notifications" or "Help me fill out this form". The agent will run commands like opencli browser <session> state or click/type. Example: open a session with opencli browser work open https://example.com. For logged-in pages, manually navigate and log in, then bind the tab with opencli browser <session> bind and let the agent drive from there.
How does this skill compare with similar options?
Compared to generic browser automation libraries like Playwright or Puppeteer, OpenCLI is designed for AI agents, providing structured output, error codes, and session management, and integrates with your own browser. It does not provide its own test runner; you'd need to integrate.