WebCLI - Browser Control for Agents

Headless bridge exposing your logged-in Chrome as browser tools for CLI AI agents like Claude Code and Codex.

WebCLI turns your Chrome into a tool your AI agent can operate. It's a headless browser-control bridge: a lightweight extension that exposes 35+ low-level browser tools — open a URL, read a page, click, type, search the web, screenshot, manage tabs, run your own JavaScript in the page — to external AI agents through a small daemon on your own machine, driven over plain HTTP. No MCP server to configure, no keys to paste. There's no agent inside the browser and no UI to babysit. You bring the model; WebCLI is the hands. Because it drives YOUR real, logged-in Chrome session, your agent works pages exactly as you do — already signed in, human-like, and sailing past the bot walls that block headless scrapers. **Reach a site nobody wrote support for.** On top of the primitives there are reconnaissance tools that answer "where does this value actually come from" — the page's structured data, its accessibility tree, and the requests it makes — plus eval_js to run your own JavaScript in the page's own origin. That is how an agent works out how to read an unfamiliar site in one session, instead of you writing and maintaining a scraper. **And page rules that stick.** Site scripts let your agent apply a change you approved — hide the clutter, restyle a page, run a small script — on every visit to a matching site, until you remove it. Every one is listed on the settings page with its matching sites and its full source, where you can pause or delete it. ━━━ WHAT YOU CAN DO ━━━ • Let Claude Code, Codex, or any agent that can run a shell command browse, read, and act on the web for you • Automate logged-in sites without pasting cookies or re-authenticating • Pull clean page text, search results, and link frontiers for research and crawling loops • Read any article as clean Markdown in a single call — no tab, no rendering, no scraping code • Fill a whole form in one call, click through flows, upload files, and drive multi-tab tasks • Work out how an unfamiliar site serves its data, then read it directly • Drive LLM-written browser tests against your real session • Fetch JSON APIs, RSS, and sitemaps straight from your session (CORS-free, cookie-authed) • Keep a page fix applied on every visit ━━━ THE TOOLBELT (35+ primitives) ━━━ • Navigate & read — open_url, get_page_text, fetch_url, get_html, screenshot, scroll_page • Search & crawl — web_search, list_links • Interact — click, type_into, fill_form, press_key, select_option, hover, drag_and_drop, file_upload, handle_dialog • Perceive the DOM — get_interactives, get_dom_outline, query_dom, find_in_page, find_in_dom, wait_for_selector • Reconnaissance — find_structured_data, get_a11y_tree, capture_network • Your own JavaScript — eval_js • Persistent page rules — create_site_script, preview_site_script, list_site_scripts, get_site_script, set_site_script_enabled, delete_site_script • Tabs — list_tabs, get_active_tab, manage_tabs, close_tab • Page-declared tools (WebMCP) — list_webmcp_tools, call_webmcp_tool ━━━ WHY WEBCLI ━━━ • Your real Chrome session — logged-in and human-like, not a throwaway headless browser • No in-browser agent — a pure browser-control provider; bring your own model and orchestration • One way in, zero config — a daemon bound to 127.0.0.1 that your agent's skill starts for you; plain HTTP, no MCP server to set up • Tiny & focused — well under a megabyte, generic tools only; runs side-by-side with the full Web Agent extension • Auto-reconnect — a 1-minute alarm redial connects within a minute of your daemon starting • Local & private — everything runs on your machine; WebCLI itself sends nothing to a third-party server (your agent, of course, sends what it reads to its own model provider) — and no web page can ever reach the tools • You hold the keys — the settings page lists every tool your agent is offered, and trims the catalog to a core set to keep its prompt small ━━━ BEFORE YOU INSTALL ━━━ • It runs headless: no window, no sidebar — a small toolbar popup for the connection, and a settings page for the tool catalog and your site scripts. • Site scripts need Chrome's "Allow user scripts" switch on this extension's details page. It takes one click, the settings page walks you through it, and every other tool works without it. ━━━ GET STARTED ━━━ 1. Install WebCLI from the Chrome Web Store (it runs headless — only a small status popup). 2. Ask your agent to run: npx skills add whitefoxx/web-tools -g 3. Then ask your agent to use the browser. The skill knows the rest — including starting the local daemon on port 9376, which WebCLI connects to automatically. Docs & the daemon: https://github.com/whitefoxx/web-tools WebCLI is the "pure provider" sibling of the Web Agent extension: the same battle-tested browser tools, packaged headless for external agents.