opencli

name	opencli
description	Use when an AI agent needs to install, inspect, troubleshoot, or operate a local OpenCLI setup, or when a task involves websites, Electron desktop apps, browser-login reuse, downloads, plugins, external CLI passthrough, or adapter generation through OpenCLI.

OpenCLI

OpenCLI turns websites, Electron apps, and existing local CLIs into structured terminal commands. Prefer it when deterministic CLI output is more useful than ad-hoc browser automation.

Trigger Cues

Load this skill when the task sounds like any of these:

• "use opencli"
• "pull data from Bilibili/Zhihu/Xiaohongshu/Weixin/Boss"
• "reuse my Chrome login"
• "control Cursor/Codex/Notion/ChatGPT from the terminal"
• "download article/video/media through OpenCLI"
• "inspect what OpenCLI supports on this machine"
• "create or troubleshoot an OpenCLI adapter"

Reference Files

Reference	Use for
references/agent-playbook.md	Agent-first trigger patterns, action order, and safe defaults
references/commands.md	Common commands, category selection, and example invocations
references/china-workflows.md	China-focused site templates and desktop-agent workflows
references/practical-scenarios.md	Detailed field manual for choosing OpenCLI and executing end-to-end workflows
references/troubleshooting.md	Failure signatures, browser bridge issues, auth problems, and fixes

Core Workflow

1. Confirm the binary exists. If opencli is missing, install it with npm install -g @jackwener/opencli.
2. Discover before guessing. Run opencli list -f yaml to inspect available adapters and commands.
3. Choose the command class:
   • Public commands: run directly. No browser login is required.
   • Browser commands: ensure Chrome is open and already logged in to the target site, then run opencli doctor if this is the first browser-backed task or if anything looks broken.
   • Desktop adapters: make sure the target app is running, inspect the registered commands first, then start with read, history, model, send, ask, or adapter-specific commands.
   • External CLI passthrough: use opencli <tool> ... and keep the original arguments intact.
   • New site development: use explore, synthesize, generate, and cascade.
4. Prefer -f json for agent-readable output. Use -f yaml for human inspection, -f table for terminal browsing, and -v only when debugging.
5. If a browser-backed command returns empty data or authorization errors, verify the Chrome login state before changing commands or code.
6. Before any write-like action such as post, reply, send, publish, greet, or follow, do at least one read-only command first unless the user explicitly asks to skip that safety step.

Decision Guide

Public data

Use this path for sources such as hackernews, wikipedia, google, arxiv, or similar read-only commands. These are the cheapest and most reliable because they do not depend on Chrome state.

Browser-backed sites

Use this path for adapters such as bilibili, zhihu, xiaohongshu, twitter, reddit, pixiv, weixin, boss, or other commands that reuse Chrome login. Keep Chrome running, keep the account logged in, and use opencli doctor before deeper debugging.

Desktop adapters

Use this path for apps such as cursor, codex, chatgpt, notion, discord-app, or doubao-app. Do not assume a command exists just because another adapter or README example has it; inspect the registry first with opencli list -f yaml or adapter help.

External CLI hub

Use this path when the user really wants the underlying CLI but wants discovery, installation help, or a unified entrypoint. Examples include opencli gh ..., opencli docker ..., opencli obsidian ..., and opencli gws ....

Adapter generation

Use opencli explore <url> --site <name> to inspect a site, opencli synthesize <site> to build an adapter, opencli generate <url> --goal "<goal>" for the fast path, and opencli cascade <url> to probe authentication strategy.

Operating Rules

• Use explicit limits like --limit 5 for exploratory reads.
• Prefer structured output over prose whenever another tool or agent will consume the result.
• For media downloads, install yt-dlp when the adapter requires it.
• When a command is likely browser-backed, assume login state matters until proven otherwise.
• For desktop adapters, check --help before assuming command parity across apps or versions.
• Treat opencli list -f yaml as the source of truth when the README and installed commands disagree.
• Read references/troubleshooting.md before concluding that an adapter is broken.