JS-View

Renders JavaScript-heavy pages in a real browser so agents can read them.

A plain fetch of a modern site usually returns an empty <div id="root"> and nothing else. The page is there, it just hasn't run its JavaScript yet. JS-View loads the URL in headless Chromium, lets the scripts run, scrolls or clicks when asked, and hands back what actually rendered as clean markdown. It plugs into Claude Code, Codex, and Cursor as an MCP server with a skill that tells the agent when to reach for it, runs as a CLI, and self-hosts as an HTTP service.

Features

• Renders SPAs, infinite-scroll feeds, and "load more" pages, then extracts the main content as markdown, text, or HTML.

• Four MCP tools: render, screenshot, links, and network (the XHR/fetch calls a page makes, with JSON bodies).

• Says when a page isn't the page: bot checks, login walls, HTTP errors, and empty renders come back labelled, with what to try next.

• Keeps agents inside their context budget: pick part of the page with a CSS selector, and long content is cut at a paragraph with the full length reported.

• Signed-in pages: sign in once in a real browser window and reuse the session in every render.

• Uses the Chrome or Edge already on the machine when Playwright's Chromium isn't installed, and keeps the browser warm between tool calls.

• Reads local dev servers from the CLI and MCP server; the HTTP service blocks private hosts by default.

• Self-hosted service with a browser pool, API-key auth, rate limiting, and Redis-backed workers for scale.

Related MCP server: fetch-mcp

Installation

JS-View needs Node.js 22 or newer. The first start downloads the package, so it can take a minute; if your agent reports that the server failed to start, run npx -y @muxover/js-view@latest --version once and restart the agent.

Claude Code — installs the MCP server and the skill:

/plugin marketplace add muxover/js-view
/plugin install js-view@js-view

Codex — add the plugin marketplace, then install js-view from /plugins:

codex plugin marketplace add muxover/js-view

Or add only the MCP server to ~/.codex/config.toml:

[mcp_servers.js-view]
command = "npx"
args = ["-y", "@muxover/js-view@latest", "mcp"]
startup_timeout_sec = 120

Cursor — one click adds the MCP server:

For the skill, copy skills/js-view into .cursor/skills/ in your project.

Any other MCP client — run the server over stdio:

{
  "mcpServers": {
    "js-view": {
      "command": "npx",
      "args": ["-y", "@muxover/js-view@latest", "mcp"]
    }
  }
}

No browser installed? JS-View uses Playwright's Chromium, or else Chrome or Edge. If none is present:

npx -y @muxover/js-view install

Quick Start

Ask your agent:

• "Read https://news.ycombinator.com and list the top five stories."

• "What API does https://example.com/products call to load its listings?"

• "Screenshot localhost:3000 and tell me what's broken in the layout."

From a terminal:

npx -y @muxover/js-view https://example.com
npx -y @muxover/js-view https://example.com/feed --scroll 3 --wait-for ".item"
npx -y @muxover/js-view localhost:3000 --screenshot home.png

Usage

MCP tools

Every tool takes url plus wait_for, wait_until, scroll, click, type, session, proxy, and timeout_ms. render adds format, selector, raw, and max_chars.

Each result starts with a header the agent reads first:

Title: Hacker News
URL: https://news.ycombinator.com/
Status: 200 | ok | 1240 ms

The status is one of ok, empty, blocked, login_required, or http_error; anything but ok comes with Note: lines on what happened and what to try.

Signed-in pages

Sign in once in a real browser window, then close it:

npx -y @muxover/js-view login https://example.com --session example

Pass session: "example" to any tool, or --session example to the CLI. Sessions are saved in ~/.js-view/sessions and updated after each render.

Commands

Flags

Content goes to stdout and the header to stderr. The exit code is 0 for a rendered page, 3 when the page was a bot check, login wall, HTTP error, or empty, 2 for bad usage, and 1 for failures.

HTTP Service

Host JS-View once and call it over HTTP, or scale renders across workers:

docker compose up --build
docker compose up --scale worker=4

Without Docker, npx -y @muxover/js-view serve starts the API on port 8080 with renders in-process. Set REDIS_URL and they go through a BullMQ queue instead.

curl -s http://localhost:8080/render \
  -H "content-type: application/json" \
  -d '{"url":"https://example.com/feed","wait_for_selector":".feed-item","actions":[{"type":"scroll","count":3}]}'

/render accepts url, output_format, wait_until, wait_for_selector, timeout_ms, actions (scroll, click, type, wait, navigate), selector, max_chars, clean, session_id, proxy, capture_network, screenshot, full_page, ocr, and screenshot_format.

Configuration

Environment variables; .env.example lists them all with comments.

Project Layout

js-view/
├── src/
│   ├── cli.ts                # js-view command: render, mcp, login, install, serve, worker
│   ├── index.ts              # Service entry for npm start and Docker
│   ├── service.ts            # HTTP API and/or worker role, graceful shutdown
│   ├── config.ts             # Environment configuration
│   ├── types.ts              # Request/response contracts
│   ├── mcp/
│   │   └── server.ts         # MCP tools: render, screenshot, links, network
│   ├── local/
│   │   ├── format.ts         # Result header for CLI and MCP output
│   │   └── login.ts          # Visible-browser sign-in that saves a session
│   ├── browser/
│   │   ├── launch.ts         # Chromium / Chrome / Edge resolution, stealth launch
│   │   ├── pool.ts           # Bounded browser pool with recycling
│   │   ├── stealth.ts        # User agent, viewport, proxy
│   │   └── session.ts        # Saved cookies and storage per session
│   ├── render/
│   │   ├── renderer.ts       # Render orchestration
│   │   ├── assess.ts         # Outcome detection and truncation
│   │   ├── actions.ts        # scroll / click / type / wait / navigate
│   │   ├── network.ts        # XHR/fetch capture
│   │   ├── screenshot.ts     # Screenshots and OCR
│   │   ├── dispatch.ts       # Inline vs queued renders
│   │   ├── tabs.ts           # Popup and new-tab tracking
│   │   └── waitStrategies.ts # Load-state and selector waits
│   ├── extract/              # Readability cleanup, markdown, text, links, metadata
│   ├── server/               # Express app, routes, validation, auth, rate limit
│   ├── queue/                # BullMQ producer and worker
│   ├── security/url.ts       # URL scheme allowlist and private-host guard
│   └── utils/                # Logger and errors
├── skills/js-view/SKILL.md   # Agent skill: when and how to use the tools
├── .claude-plugin/           # Claude Code plugin and marketplace
├── .agents/plugins/          # Codex plugin marketplace
├── plugin.json, mcp.json     # Agent Plugins manifest and MCP server
├── .mcp.json                 # MCP server for the Claude Code plugin
├── tests/                    # Unit, browser, MCP, and CLI tests
├── docs/PROJECT.md           # Maintainer notes
├── Dockerfile                # Playwright-based service image
└── docker-compose.yml        # API, worker, and Redis

Limitations

• Sites with strong bot protection can still block it. JS-View reports the block rather than returning the challenge page as content.

• A render takes one to a few seconds; heavy pages take longer.

• OCR is best-effort, for text baked into images.

• Signing in needs a desktop with a display, since the login window is a real browser.

Contributing

See CONTRIBUTING.md.

License

Licensed under the MIT license.

Links

• Repository: https://github.com/muxover/js-view

• Issues: https://github.com/muxover/js-view/issues

• Changelog: CHANGELOG.md

• npm: https://www.npmjs.com/package/@muxover/js-view