mcp-fetch-ux

MCP server that fetches web pages using a real, stealth browser. Renders JavaScript, pierces Shadow DOM, returns clean text — and gets through bot walls that block everything else, including live Google Search results. Can interact with pages — click buttons, fill forms, download files.

Built because Claude Code's WebFetch hangs indefinitely on slow sites with no timeout, and existing fetch tools can't see inside Shadow DOM (or get captcha'd the moment they try).

How it works

1. A stealth browser launches headed under xvfb and navigates to the URL — by default a fingerprint-patched Firefox that passes reCAPTCHA; swappable to real Chrome (see Browser engine)

2. Dismisses cookie/consent overlays automatically (including late, JS-injected banners)

3. Waits for JS to render (polls until page content stabilizes)

4. Ctrl+A, Ctrl+C — selects and copies the rendered page, capturing all visible text including content inside (closed) Shadow DOM

5. Discovers available actions (buttons, links, inputs) and returns them as hints

6. Optionally runs actions (click, fill, wait) — if a click triggers a download, returns the file content

No LLM in the loop. No API costs. 30-second timeout.

Related MCP server: MCPBrowser

Quick start

git clone https://github.com/bxxd/mcp-fetch-ux.git
cd mcp-fetch-ux
make install   # deps + the invisible engine (stealth Firefox) + 'fetch' CLI to ~/.local/bin/
# (optional) the chrome engine too:  make chrome

# Basic fetch
fetch https://www.roche.com/solutions/pipeline

# Click a button to download CSV
fetch https://roche.com/solutions/pipeline --click "button:has-text('Download current view as CSV')"

# Save to file
fetch https://roche.com/solutions/pipeline --click "button:has-text('Download current view as CSV')" -o pipeline.csv

# More content (default 50K chars via MCP, unlimited via CLI)
fetch https://en.wikipedia.org/wiki/Likelihood_ratio --max 100000

Start the MCP server:

make server    # start on port 5006

MCP configuration

{
  "mcpServers": {
    "fetch-ux": {
      "type": "sse",
      "url": "http://127.0.0.1:5006/sse"
    }
  }
}

Tool

fetch

Two-step interaction

First call returns page content + available actions:

Contents of https://www.roche.com/solutions/pipeline:
Title: Roche | Product Development Pipeline

RG7716
faricimab
Vabysmo
macular edema secondary to branch retinal vein occlusion (BRVO)
...

---
Available actions on this page:
  - click: "button:has-text('Download current view as CSV')"
  - click: "button:has-text('Phase')"
  - fill: "search" (search)

Second call with actions gets the data:

{
  "url": "https://www.roche.com/solutions/pipeline",
  "actions": [{"action": "click", "selector": "button:has-text('Download current view as CSV')"}]
}

Returns the full CSV (51K chars, 131 pipeline entries with descriptions).

Why not just httpx/curl?

They don't render JavaScript. Roche's pipeline page returns an empty shell — the drug data loads via JS into Shadow DOM web components. curl gets nothing. This tool gets what a human sees.

Why not innerText or Readability?

page.innerText('body') and document.getSelection() don't cross Shadow DOM boundaries. Readability can't see JS-rendered content. The clipboard route captures exactly what a user gets when they Ctrl+A, Ctrl+C in a real browser — the only reliable way to get all visible text (including closed Shadow DOM) from modern web pages.

Why not crawl4ai?

Tested crawl4ai (50K+ stars) on the same Roche pipeline page. It returns 7,220 chars with zero drug names — can't see inside Shadow DOM. This tool returns 11,224 chars with all 131 drugs.

Browser engine

The browser is a swappable engine — pick with FETCH_UX_ENGINE:

Both run a real browser headed, so they need a display — run under xvfb (the systemd unit and cli already do). make setup installs only the default (invisible) engine's browser; make chrome adds Chrome.

GPU: with a DRM render node (/dev/dri/renderD128) present — e.g. a GPU passed into the container — the chrome engine drives WebGL through it via ANGLE/EGL, so the renderer reports the real GPU instead of WebGL: false. No-op without a GPU.

Performance

The browser launches once at server startup and stays warm; each fetch opens and closes a page.

Concurrency is per-engine: invisible runs fetches one at a time (a single Firefox can't safely open targets in parallel), chrome runs up to 3.

Dependencies

• Engines — invisible_playwright (stealth Firefox, default) · Patchright (stealth Chrome)

• Playwright — browser automation

• Starlette + Uvicorn — HTTP/SSE server

• MCP SDK — Model Context Protocol

License

MIT