An MCP server gives an AI assistant tools it can call. With a Markdown server connected, you can ask Claude, ChatGPT, Cursor, or Codex to read a page, and it gets the text as Markdown instead of HTML: headings, lists, links, and tables, without scripts and markup. That is fewer tokens and less noise for the model.
I’m the founder of ScreenshotOne, which is the first server on this list. The sections on the others are based on their own READMEs, documentation, and pricing pages as of November 16, 2026.
Local servers, hosted servers, and JavaScript
Two things decide which server fits.
Where it runs. A local server is a program on your computer that your assistant starts. Only assistants that run on your computer can do that: Claude Desktop, Claude Code, Cursor, VS Code, Codex, and similar. A hosted server runs on the vendor’s side, and your assistant connects to its URL, so it also works in Claude on the web and in ChatGPT, and nobody installs anything.
How it gets the page. Some servers download the HTML and convert it. That is fast and cheap, and it works for articles and documentation. But many sites build the page with JavaScript, and the HTML they send is almost empty. For those, the server has to open the page in a real browser and convert what the browser rendered.
Comparison
| Server | Where it runs | Sign-in | Renders JavaScript | Besides Markdown | Price |
|---|---|---|---|---|---|
| ScreenshotOne | Hosted, or local | OAuth | Yes | Screenshots, PDFs, HTML to image | 100 free a month, then from $17 a month |
| Firecrawl | Hosted, local, or self-hosted | None, OAuth, or API key | Yes | Crawl, search, documents | 1,000 free credits a month, then from $19 a month |
| Jina AI | Hosted | None, or API key | Yes | Search, PDFs, screenshots | Free at 20 requests a minute |
| Fetch | Local, open source | None | No | — | Free |
| MarkItDown | Local, open source | None | Not stated | PDFs and Office files | Free |
| Tavily | Hosted, or local | OAuth or API key | Not stated | Search, crawl, map | 1,000 free credits a month, then from $30 a month |
| Exa | Hosted | None, OAuth, or API key | Not stated | Search | $10 of free credit a month |
| Bright Data | Hosted, or local | API token | No, in its Markdown tool | Search, site scrapers, remote browser | 5,000 free requests a month |
| Cloudflare Browser Run | Hosted | Cloudflare OAuth | Yes | Screenshots, PDFs, crawl | 10 free browser minutes a day |
| Crawl4AI | Self-hosted, open source | Optional JWT | Yes | Screenshots, PDFs, crawl | Free |
1. ScreenshotOne
ScreenshotOne’s MCP server is a hosted server. You add https://mcp.screenshotone.com to your assistant and sign in to ScreenshotOne in your browser. There’s no API key to copy into a config file, and you can disconnect the assistant at any time. There are setup guides for Claude, Claude Code, ChatGPT, Codex, Cursor, VS Code, and other clients. If you’d rather run it locally, the same tools are in an open-source server.
The extract-website-markdown tool takes a URL, renders the page in a browser, and returns the cleaned Markdown as text. Because the page is rendered, it works on sites built with JavaScript.
What makes it different from the rest of this list is the screenshot. An agent that reads a pricing page as Markdown knows what it says, but not how it looks. The render-website-screenshot tool can return a screenshot and the Markdown from the same render, without ads and cookie banners, and split a long page into slices the model can read. You might ask:
Read https://example.com/pricing as Markdown, take a full-page screenshot,and tell me which plan the page pushes visitors toward.It doesn’t crawl a site or search the web, the Markdown tool has no options besides the URL, and it reads public pages only. If you need to read a whole site, look at Firecrawl or Crawl4AI below.
The free plan has 100 screenshots a month without a credit card, and paid plans start at $17 a month. MCP calls use the same credits as the API.
2. Firecrawl
Firecrawl has the most complete server for reading the web. It’s hosted at https://mcp.firecrawl.dev/v2/mcp, and it also runs locally with npx or self-hosted. Its firecrawl_scrape, firecrawl_search, and firecrawl_parse tools work without a key at a limited rate; the rest need OAuth or an API key. firecrawl_scrape returns Markdown by default, can keep only the main content, and Firecrawl’s docs say it handles JavaScript-rendered sites, PDFs, and proxies. It also crawls and maps whole sites, searches the web, and parses PDF, Word, and spreadsheet files. The full profile has 33 tools. The free plan has 1,000 credits a month, and Hobby is $19 a month for 5,000 credits, with one credit per scraped page.
3. Jina AI
Jina AI’s MCP server is hosted at https://mcp.jina.ai/v1. Its read_url tool reads a web page or a PDF as Markdown, up to five URLs in a call, and works without a key at a lower rate limit (Reader allows 20 requests a minute without one). Jina says Reader renders the page in a headless browser by default. The server also has web, arXiv, and image search and a screenshot tool, but it doesn’t crawl sites. Its README notes that the output is cut to fit a 25,000-token limit unless you change it, and that a blocked page comes back as its login wall, with no error.
4. Fetch
Fetch is the reference server from the MCP project: free, MIT-licensed, and local only. Its fetch tool downloads the page and converts it to Markdown without a browser, so a page built with JavaScript can come back nearly empty. It returns 5,000 characters by default, and the model asks for the next part with start_index. It obeys robots.txt when the model starts the request. For documentation and articles on your own machine, it’s all you need.
5. MarkItDown
MarkItDown is Microsoft’s converter, with one tool, convert_to_markdown, that takes an http:, https:, file:, or data: URI. It’s best at files: PDFs, Word, Excel, and PowerPoint documents. Its README says it’s meant for local use, it has no authentication, and the package is still an alpha release. It doesn’t say whether it renders JavaScript.
6. Tavily
Tavily is a search API for agents. It’s hosted at https://mcp.tavily.com/mcp with OAuth or an API key, or runs locally. Its tavily-extract tool returns Markdown by default for up to 20 URLs in a call, and there are search, map, and crawl tools next to it. The free plan has 1,000 credits a month, and a basic extract costs one credit per five URLs. Paid plans start at $30 a month.
7. Exa
Exa is also search first. It’s hosted at https://mcp.exa.ai/mcp, works without signing in at a limited rate, and has OAuth and API keys for more. web_fetch_exa reads one or more pages as clean Markdown; its parameters aren’t documented. Exa gives $10 of free credit a month, and page contents cost $1 per 1,000 pages.
8. Bright Data
Bright Data’s server is hosted, with an API token in the URL, or local. scrape_as_markdown goes through Bright Data’s unblocking proxies, which helps with sites that block other servers on this list. Its README says the scrape tools don’t render JavaScript; for those pages there’s a separate set of remote browser tools that return an accessibility snapshot instead of Markdown. Pages are cut at 40,000 characters, and the model can ask for the rest. The free tier has 5,000 requests a month without a card.
9. Cloudflare Browser Run
Cloudflare’s Browser Run server (the product used to be called Browser Rendering) is hosted at https://browser.mcp.cloudflare.com/mcp and signs in with your Cloudflare account. get_url_markdown renders the page in a real browser, and there are screenshot, PDF, link, and crawl tools as well. Cloudflare’s docs say its requests are always identified as a bot, so protected sites can refuse them. The Workers free plan includes 10 browser minutes a day.
10. Crawl4AI
Crawl4AI is an open-source crawler (Apache-2.0) whose Docker server has MCP endpoints built in. You host it yourself, so there’s no vendor URL to connect to, and it costs only what your server costs. Its md tool returns Markdown, and it has html, screenshot, pdf, execute_js, and crawl tools.
Servers you’ll see recommended that don’t fit
- Playwright MCP and Chrome DevTools MCP return an accessibility snapshot of the page, not Markdown. That’s the right format for clicking through a page, and a noisy one for reading an article. See the Playwright MCP and Chrome DevTools MCP guides for what they’re good at.
- The Puppeteer reference server was moved to the archived servers repository in May 2025.
- ScrapingBee’s hosted server lists tools for a page’s HTML, a screenshot, and a file on its own page, but not a Markdown tool.
Apify, Scrapfly, Scrapeless, and Spider also have hosted servers that can return page content. They’re scraping platforms first, and worth a look if you already use one of them.
Which one should you choose?
Use ScreenshotOne when your agent needs to see the page as well as read it: reviewing landing pages, comparing pricing pages, checking a release, or attaching a screenshot to a report. One connection gives it the Markdown and the screenshot.
The others make sense in these cases:
- Reading a whole site, or searching and reading in one server: Firecrawl.
- A page now and then, with no account: Jina AI, or Fetch if you work in a desktop or command-line client.
- PDFs and Office files on your disk: MarkItDown.
- Research that starts with a search query, not a URL: Tavily or Exa.
- Sites that block ordinary requests: Bright Data.
- You already have a Cloudflare account: Cloudflare Browser Run.
- Everything on your own infrastructure: Crawl4AI.
To try ScreenshotOne, add https://mcp.screenshotone.com to your assistant and sign in. If you don’t have an account yet, you can create one for free. The setup guide has the steps for each assistant. If screenshots are the main job, see the comparison of screenshot MCP servers, or the comparison of the best screenshot APIs if your own code requests them.
Frequently Asked Questions
If you read the article, but still have questions. Please, check the most frequently asked. And if you still have questions, feel free reach out at support@screenshotone.com.
What is a Markdown MCP server?
An MCP server with a tool that takes a URL and returns the page as Markdown. An AI assistant such as Claude, ChatGPT, Cursor, or Codex calls it to read a page without the navigation, scripts, and markup of the HTML.
Which Markdown MCP servers work in Claude on the web and in ChatGPT?
Hosted ones: ScreenshotOne, Firecrawl, Jina AI, Tavily, Exa, Bright Data, and Cloudflare Browser Run all have a remote URL. Fetch and MarkItDown run only on your computer, and Crawl4AI is self-hosted.
Which Markdown MCP servers render JavaScript?
ScreenshotOne, Jina AI, Cloudflare Browser Run, and Crawl4AI render the page in a browser, and Firecrawl's docs say it handles JavaScript-rendered sites. Fetch downloads the HTML without a browser, and Bright Data's README says its scrape tools don't render JavaScript.


