OpenCLI is a browser automation CLI for humans and AI agents that works through your logged-in Chrome session. It turns websites into deterministic commands through site adapters, and it also lets an agent navigate, inspect, click, fill forms, extract, and screenshot any page without an adapter. It presents itself as an alternative to Agent-browser.
It covers three kinds of automation. Built-in adapters expose reusable commands for more than 100 sites, including Bilibili, Zhihu, Xiaohongshu, Reddit, Hacker News, Twitter/X, LinkedIn, Amazon, and web front ends for Claude, Gemini, and NotebookLM. Agent skills, installed into an agent such as Claude Code or Cursor, let the agent operate any website on your behalf. An adapter-authoring skill guides an agent through writing a new adapter, from site reconnaissance and endpoint discovery through authentication choice, field decoding, and verification, and keeps site knowledge for the next adapter on the same site.
OpenCLI connects to Chrome or Chromium through a Browser Bridge extension and a small local daemon, and it can also attach to a remote browser over the Chrome DevTools Protocol. Agents read pages through structured DOM snapshots rather than screenshots, and can intercept network API responses. A desktop app for macOS and Windows bundles the runtime with a tray UI for setup, diagnostics, updates, keeping browser logins alive, and converting web pages to Markdown.
Commands output tables, JSON, YAML, Markdown, or CSV, and follow Unix exit codes so scripts and CI can branch on failure modes such as an empty result, a disconnected bridge, or missing authentication. Community plugins add further adapters, and you can eject and modify official adapters locally.
Features
- Site adapters: deterministic commands for more than 100 sites, from Hacker News and Reddit to Bilibili, Xiaohongshu, and LinkedIn
- Agent browser control: agents navigate, read, click, fill, extract, and wait on any page through your logged-in Chrome
- Agent skills: skills for browser control, adapter authoring, adapter repair, and a command reference
- Adapter authoring: a guided flow from site reconnaissance to a verified adapter, with site knowledge kept for reuse
- DOM snapshots: pages are read as structured DOM rather than screenshots, with network response interception
- Chrome profiles: routes commands to named Chrome profiles when several are connected
- Remote browsers: attaches to a remote browser over the Chrome DevTools Protocol
- Downloads: images, videos, audio, transcripts, and articles from supported platforms
- Output formats: table, JSON, YAML, Markdown, and CSV
- Plugins: install, update, and publish community adapters from GitHub
- Desktop app: a tray app for setup, diagnostics, updates, login keepalive, and web-to-Markdown conversion
