Surfboard

The Web Browser for LLMs

Most LLM "search" tools are really just glorified web scrapers.
Surfboard lets LLMs click links, scroll, fill forms, switch tabs, and navigate websites through a token-efficient MCP server.
Surfboard uses Playwright (headless Chromium) to render JS-heavy pages -- sidestepping many bot walls, empty JS shells, and client-side rendering.
By structuring results as tree-like JSONs, Surfboard both prevents token overflows on large sites and also allows the LLM to navigate intentionally and efficiently.

Install

After cloning the repo, install the MCP tool using pip

pip install -e .

If your AI tool supports auto discovery (e.g. Claude Desktop, Cursor, OpenCode, VS Code Copilot, Windsurf, Claude Code CLI), run:

surfboard setup

This registers Surfboard as an MCP server in your tool's config. After restarting the tool, the LLMs should be able to browse and interact with the web more thoroughly using Surfboard.

CLI commands

Typical workflow

1. Search — surfboard_search("query") runs a Google search
2. Navigate — surfboard_browse(url) opens a page, returns structured sections + interactive elements
3. Read — surfboard_get_page returns the page structure with sections, links, buttons, inputs
4. Dig deeper — surfboard_get_full_text grabs document.body.innerText for JS-rendered content; surfboard_evaluate("JS code") runs arbitrary JavaScript
5. Interact — surfboard_click(id), surfboard_fill(id, text), surfboard_fill_and_submit(id, text), surfboard_hover(id), surfboard_scroll_to(id), surfboard_scroll_by(x, y), surfboard_wait_for_load(timeout_ms), surfboard_press_key(key)

MCP tools

surfboard_search surfboard_browse surfboard_click surfboard_fill surfboard_hover surfboard_scroll_to surfboard_scroll_by surfboard_wait_for_load surfboard_get_page surfboard_back surfboard_forward surfboard_tab_new surfboard_tab_switch surfboard_refresh surfboard_status surfboard_evaluate surfboard_get_full_text surfboard_screenshot surfboard_fill_and_submit surfboard_press_key surfboard_clipboard_copy surfboard_clipboard_read surfboard_highlight

Disclaimer

Surfboard is designed for interaction, not bulk extraction. If an LLM only needs a one-shot context dump from a webpage, a traditional scraper is usually faster. When it needs to navigate, click, search within a site, or interact with dynamic content, Surfboard is superior.

Tricks & anti-blocking

• Reddit URLs auto-rewrite to the JSON API (no CAPTCHA)
• Twitter/X URLs auto-rewrite to Nitter (no JS wall)
• Medium URLs auto-rewrite to Scribe (no paywall)
• AMP URLs auto-strip the /amp suffix
• DuckDuckGo/Google redirect links are unwrapped to real URLs
• Bot-block patterns (Cloudflare, CAPTCHA, etc.) are detected and reported