cloudflare browser runsnapshotmarkdownllmweb2md

Cloudflare Browser Run /snapshot vs a Clipper

Zephyr Whimsy2026-11-219 min read

On 19 November 2026, Cloudflare's Browser Run changelog expanded /snapshot so one page load can also return pdf, links, scrape, and json, alongside HTML (content), a screenshot, Markdown, and the accessibility tree. Use that call when one rendered page must come back as several formats. When an LLM, a RAG chunk, or an agent needs one live page as clean Markdown, call /markdown, or clip the tab you already have open.

Checked on 21 November 2026 (Asia/Shanghai): the 19 November changelog, the /snapshot and /markdown quick actions, and the snapshot API reference. The Web2MD homepage and the other Web2MD pages linked here were re-read the same day.

Short answer

Use /snapshot when one Browser Run page load must return at least two formats, including the pdf, links, scrape, and json formats added on 19 November 2026. Use /markdown, or a clipper on the open tab, when an LLM only needs that page as Markdown.

What did Browser Run add to /snapshot on 19 November 2026?

The changelog is titled "Capture PDFs, links, scraped elements, and extracted JSON with /snapshot". Before it, /snapshot returned HTML, screenshots, Markdown, and the accessibility tree, so a PDF or structured data meant loading the page again.

pdf prints the page, with the same pdfOptions as /pdf. links collects links, with the same visibleLinksOnly and excludeExternalLinks options as /links. scrape returns the elements matched by elements, the same payload as /scrape, and elements is required. json extracts structured data from the page HTML with Workers AI, with the same prompt, response_format, and custom_ai fields as /json. The changelog says prompt or response_format is required, and the request incurs Workers AI usage.

Omit formats and the default is ["content", "screenshot"]. The Accept-header feature is a different product: Cloudflare Markdown for Agents.

Which formats can one /snapshot call return?

Eight values are accepted. You must request at least two distinct formats. A single format uses its own endpoint. For Markdown, that endpoint is /markdown.

| Format | Result field | Attached options | Single-format endpoint | | --- | --- | --- | --- | | content | result.content | HTML | /content | | screenshot | result.screenshot | screenshotOptions, selector, scrollPage | /screenshot | | markdown | result.markdown | Markdown | /markdown | | accessibilityTree | result.accessibilityTree | interestingOnly, root | /accessibilityTree | | pdf | result.pdf | pdfOptions | /pdf | | links | result.links | visibleLinksOnly, excludeExternalLinks | /links | | scrape | result.scrape | elements required | /scrape | | json | result.json | Workers AI | /json |

The JSON requirement disagrees across pages checked on 21 November 2026. The changelog says prompt or response_format is required. The /snapshot table marks prompt as required and also lists response_format and custom_ai. The API reference marks prompt and response_format as optional schema fields. Use the page you are coding against.

json incurs Workers AI usage. One actionTimeout covers navigation, captures, retries, and the AI request: default 60 seconds, maximum 120 seconds. No page checked here publishes a Workers AI price, a rate limit, or a plan name, so none is added.

Formats are captured in a fixed order, not array order. The changelog says the PDF is printed last because print handlers can change the page. A live page can change between captures. If any format fails, the whole request fails.

Send url or html. The quick-action page shows POST https://api.cloudflare.com/client/v4/accounts/<accountId>/browser-run/snapshot with a Browser Run - Edit token, or env.BROWSER.quickAction("snapshot", ...) from a Worker with no token. The API reference titles it POST /accounts/{account_id}/browser-rendering/snapshot. Those path segments disagree, and this article does not pick one.

When is /markdown the right Browser Run call?

/markdown turns a webpage into Markdown for summaries, diffs, embeddings, or a saved article with the styling and scripts stripped. The path is POST https://api.cloudflare.com/client/v4/accounts/<accountId>/browser-run/markdown, or env.BROWSER.quickAction("markdown", { url }), with url or html.

For JavaScript-heavy pages, both docs say the default load can be empty because the browser finishes before scripts do. Both point to gotoOptions.waitUntil of networkidle0 or networkidle2, and to waitForSelector when you know the element.

Both allow userAgent. Both say it does not bypass bot protection, that Browser Run requests are always identified as a bot, and that servers should use the non-configurable headers rather than the User-Agent string.

Use /markdown for one Markdown document. Use /snapshot when that load must also return another accepted format. A formats list of only "markdown" breaks the two-format rule.

When is a Markdown clipper the better handoff to an LLM?

Browser Run fetches a URL as a bot. A clipper reads the document already rendered in your browser. That matters when a login, a challenge, or the tab's own scripts produced the text.

On 21 November 2026 the homepage says install the extension, open the page, click the icon, and copy Markdown to an AI tool. Docs say Alt+M or the icon. llms.txt still says "Last verified: 2026-11-14" and lists Ctrl+M / Alt+M. The homepage and pricing FAQ name Chrome and Firefox, with Edge in development and Safari planned. The docs introduction names Chrome only.

Send to AI on the homepage targets Claude, ChatGPT, Gemini, Perplexity, Grok, or DeepSeek. That page says 3 free conversions a day and no daily cap on Pro. Pricing lists Free at $0 (3 a day, latest 50 library items), Pro Monthly at $9/month under a struck $15 and "40% OFF", Yearly at $49.99/year ("Just $4.17/month"), a 30-Day Pro Pass at $15, and Lifetime at $149. The homepage Pro cell shows $4.17/mo beside a struck $15: the yearly line, not the $9 price.

The homepage FAQ and the pricing FAQ say every conversion stays local and is never uploaded. llms.txt says the free tier converts public pages on Web2MD's server and Pro converts in the signed-in browser. The changelog treats saving as something you can turn off: v1.8.0 (17 August 2026) adds Privacy Mode, and v2.0.0 (26 October 2026) adds chat export for ChatGPT, Claude, Gemini, DeepSeek, and Perplexity, plus a browser-made PDF it calls unlimited and free.

Agent Bridge is the signed-in path. It says use a hosted URL-to-Markdown API for ordinary public URLs, and the bridge after login, a challenge, or JavaScript. It uses your Chrome session and does not bypass authentication, permissions, MFA, or a CAPTCHA. Chrome and the extension must be running. agent_convert takes one URL. agent_batch_convert takes up to 50. Setup is npm i -g web2md-mcp-server && web2md-mcp-setup. Changelog v1.9.3 (5 September 2026) names Claude Code, Codex, Cursor, and Claude Desktop.

The homepage's own figures are about 65% fewer tokens than raw HTML, and 1,000+ installs. llms.txt says the benchmark does not report a token percentage.

| | /snapshot | /markdown | Clipper on the open tab | | --- | --- | --- | --- | | Returns | Two or more formats from one load | Markdown of a URL or HTML | Markdown of the rendered tab | | Use for an LLM | You also need another format from that load | One Markdown page from a bot-identified load | The open tab, including a session the bot lacks | | Session | Always a bot. userAgent does not bypass protection | Same on the /markdown page | Your browser. Agent Bridge keeps your Chrome login and does not bypass it | | Rule | Two formats minimum. elements for scrape. JSON uses Workers AI | Not a one-item /snapshot | Browser and extension stay open. Unloaded items can be missing | | Failure | One failed format fails the request. PDF prints last | That single call fails or returns Markdown | You keep what has already rendered |

How do you hand one live page to an LLM, a RAG index, or an agent?

  1. Put the URL and the capture day at the top of the Markdown you keep.
  2. For Markdown only, call /markdown with url or html. Do not send /snapshot one format.
  3. For a second format from that load, call /snapshot with at least two formats. Add elements for scrape. For json, the changelog wants prompt or response_format, and the table marks prompt required. Expect Workers AI usage. PDF prints last. One failure fails the request.
  4. If you are looking at the tab, or it rendered only because you are signed in, clip it: icon, then copy or Send to AI. An agent keeps Chrome up and calls agent_convert or agent_batch_convert.
  5. Use /snapshot for a bundle, such as PDF plus links. Use the clipper for the page a model should quote.

Check pricing again before you pick a plan. The Cloudflare pages opened on 21 November 2026 do not list a Browser Run price or rate limit.

How is this different from Cloudflare Markdown for Agents?

Cloudflare Markdown for Agents is content negotiation: the client sends Accept: text/markdown to a zone the owner enabled. You cannot turn that on for someone else's article.

/snapshot and /markdown load a url or html in Browser Run and return the formats you named. A clipper reads a tab that has already rendered. Use the earlier post for the Accept header, and this one to choose the snapshot, /markdown, or the open page.

Which snapshot and clipper facts should you recheck?

What did /snapshot add on 19 November 2026?

pdf, links, scrape, and json, on top of HTML (content), a screenshot, Markdown, and the accessibility tree.

Can /snapshot return only Markdown?

No. It needs at least two distinct formats. Markdown alone is /markdown.

What do scrape and json require?

scrape needs elements. The changelog says json needs prompt or response_format. The table marks prompt required. json uses Workers AI. No usage price is published on the pages checked.

What happens if one format fails, and when is the PDF taken?

One failure fails the whole request. Order is fixed. The changelog prints the PDF last because print handlers can change the page, and the live page can change between captures.

Does userAgent make Browser Run look like a normal browser?

No. Both pages say userAgent does not bypass bot protection, and Browser Run is always identified as a bot.

When is a clipper the better tool?

When you need the open tab as Markdown, including a login the bot will not carry. Agent Bridge uses that Chrome session and does not bypass it. Public Markdown-only URLs use /markdown. Two or more formats use /snapshot.

Related Articles

Most Read

last 30 days
  1. #1Can Claude Read Reddit? No — Here's the One-Click Fix (2026)
  2. #2HTML vs Markdown for LLMs: I Wasted 67% of My Tokens for a Year
  3. #3Reducing Token Waste in ChatGPT and Claude: 7 Techniques That Cut Costs 72%
  4. #4Obsidian Web Clipper Official Plugin 2026: Complete Guide + When You Need More
  5. #5Reddit JSON API vs Scraping: The Honest 2026 Comparison for Developers

Latest Articles