Supported sources

anymd picks a converter per URL. Specialised adapters go first; everything else goes through the general web pipeline. The kind field in the frontmatter (and the X-Anymd-Kind header) tells you which one ran.

At a glance

Source	kind	What you get	Credits
Any web page	web	Main content with clutter removed	1
X / Twitter	x	Posts, long-form Articles, quotes, polls, media, engagement stats	1
Facebook	facebook	Public post text, author, media and optional comments	10
Instagram	instagram	Public post or reel text, author and media; optional comments	10
Threads	threads	Public post text, author and carousel media; optional comments	10
LinkedIn	linkedin	Public post text, author, media and optional comments	10
YouTube	youtube	Title, channel, thumbnail and a timestamped transcript	3
GitHub	github	READMEs, issues, PRs, discussions	1
Reddit	reddit	Threads with comments	1
Hacker News	hackernews	Story plus the top of the discussion, nested	1
PDF	pdf	Text with layout and tables kept	3
DOCX, XLSX, XLS, ODS, ODT, CSV	document	Documents as Markdown, sheets as tables	3
Images (JPEG, PNG, WebP, SVG)	image	A vision-model description and transcription	5
Plain text, Markdown, JSON URLs	text	Passed through (JSON in a code block)	1

Cached results cost 0. Failed conversions cost 0. Full pricing: Billing & credits.

Web pages

The web pipeline fetches the page, parses it with linkedom, and runs a content-extraction engine that finds the main content, removes navigation, ads and boilerplate, standardises headings, code blocks and footnotes, and emits Markdown.

anymd adds a few things on top:

• Retry for empty shells. If a client-rendered page comes back with almost no text, anymd retries once with its bot user agent and keeps the richer result. It always identifies as anymd.
• GitHub gets a bot user agent first, because it serves server-rendered content to bots.
• Documents behind URLs. If a URL serves a PDF, Office file or image, it is routed to the document converter automatically.
• Your overrides. selector picks the content block when automatic detection misses; images=0 strips images; lang sets Accept-Language. See URL API.

Limit: 5 MB of HTML per page.

Site rules anymd follows

Before fetching a page itself, anymd checks the site's robots.txt for the anymd token (or *) and refuses disallowed pages with 403 robots_disallowed. If robots.txt cannot be read because of a server error, the request fails with 503 robots_unreachable. Each site also has a shared per-minute fetch budget (429 domain_rate_limited), and domains whose owners opted out are refused on every source with 403 site_opted_out. None of these use credits. Site owners: see Site Owners & Abuse.

X / Twitter

Status URLs on x.com and twitter.com (including mobile.) go through the FxTwitter API. You get the post text, long-form Articles, quoted posts, polls, photos and videos, plus likes, retweets, replies and views in the frontmatter.

Every read converts the requested post. Signed-in callers can opt in to the rooted same-author thread with expandThread (per request or as a saved default), bounded by maxThreadPosts (default 20, maximum 100); each additional post is a separate 1-credit unit. Threads no longer expand automatically for signed-in callers. Thread membership follows the provider's relationship and conversation identifiers, so unrelated replies and other authors are excluded.

Social posts

Facebook, Instagram, Threads and LinkedIn adapters require an account and a configured RAPIDAPI_KEY; they accept public post URLs and cost 10 credits for the base post. Comments are opt-in with includeComments=1 and are charged in started batches of 20. The response can include media without enabling comment retrieval.

Comment retrieval depends on upstream provider availability. Threads uses the threads-api4.p.rapidapi.com provider and supports nested replies within the bounded request budget. When that provider reports has_next_page: true without a cursor, anymd marks coverage source_incomplete instead of treating the page as a successful complete result. Instagram base post, media and image analysis are available, but chronological comments can time out and fall back to a popular subset; reply requests can return 429, so reply coverage is not fully proven. Check enrichment.comments.complete, count and reason instead of assuming comments are complete.

YouTube

Watch, youtu.be, Shorts and Live URLs. Title, channel and thumbnail come from YouTube's oEmbed endpoint. The transcript comes from a third-party caption provider and is grouped into readable paragraphs with timestamps:

## Transcript

**0:00** · Welcome back. Today we're looking at…

**0:42** · The first thing to notice is…

Pass lang=vi (or any language code) to prefer a transcript language. When no transcript is available, the output says so and still includes the video's metadata. That conversion is still charged.

Video download to the anymd CDN

Off by default, account required. Ask for it on one read with downloadVideo: true (REST and MCP), ?downloadVideo=1 (URL API) or --download-video (CLI), or turn on Download YouTube videos to the anymd CDN in Account → Deep reading defaults (or PUT /api/v1/account/reading-preferences with {"downloadVideo": true}) for every YouTube read; downloadVideo: false on a request switches the saved default off once. The read then also downloads the lowest-quality MP4 of the video (the smallest stream that has sound; a video-only stream only when no other exists, up to 500 MB) to cdn.anymd.cc, in the background. The read does not wait for it: the Markdown ends with a section like this, and the JSON response carries the same data in video_download:

## Video download

The lowest-quality MP4 of this video is being downloaded to the anymd CDN in the background.

- Job: `vid_…` (status: `queued`)
- Check: `GET https://anymd.cc/api/v1/videos/vid_…` with your API key, or the MCP tool `get_video_download` with `{"id": "vid_…"}`.
- Poll every 15 to 30 seconds until `status` is `ready` (the MP4 is at `cdn_url`) or `failed` (see `error`). 20 credits are charged only when it is ready.

Status goes queued → downloading → ready or failed. Reading the same video again returns the running or finished job instead of starting a new one. When the account cannot pay for a download or the feature is unavailable, the section says why and no job starts.

AI video analysis

Also off by default. Per read, analyzeVideo: true (?analyzeVideo=1, --analyze-video) asks for the download and its analysis together; as a saved default, turn on Analyze downloaded videos with AI ({"analyzeVideo": true}) next to video download. Once the MP4 is ready, Google's Gemini 3.8 Flash (google/gemini-3.8-flash, through OpenRouter) watches and listens to it and writes a Markdown analysis: summary, a timeline with mm:ss timestamps, spoken content, on-screen text and visual details, in the video's main language. Keep polling the same job until analysis.status is ready (the result is analysis.markdown) or failed (see analysis.error). A later read of a video whose analysis is ready includes it in the Markdown under Video analysis (AI-generated).

It costs 10 credits plus 20 per started minute of video (a 19-second clip is 30 credits, a 3.5-minute video 90), charged only when the analysis is ready. Videos longer than 60 minutes are not analyzed (video_too_long), and an account without enough credits for the video's length gets quota_exceeded before the model runs. Turning the setting on later and reading a video again adds an analysis to its existing download.

Hacker News

news.ycombinator.com/item?id=… uses the official Hacker News API: the story (points, author, comment count, link, text) and the top 20 comment threads, each with up to 3 replies, nested as blockquotes up to 3 levels deep.

Files and documents

PDFs, images, Office documents and spreadsheets are converted with Cloudflare Workers AI toMarkdown:

• PDF: text layout kept, tables preserved.
• DOCX, ODT: document text as Markdown.
• XLSX, XLS, ODS, CSV: sheets become Markdown tables.
• Images: a vision model describes and transcribes the image; the Markdown starts with the image itself.

Convert a file behind a URL with the URL API, or upload one with POST /api/v1/convert/file or anymd file <path>. Limit: 20 MB per file.

What won't convert

• Private networks, localhost, internal hostnames and URLs with credentials. They're blocked on purpose.
• Pages behind a login or paywall. anymd sees what an anonymous visitor sees.
• Content types other than HTML, text, JSON and the document types above return 415 unsupported_type.

If a public page converts badly, open an issue with the URL. That's how adapters get better.

Planned

These are on the roadmap and not available yet:

• Audio and podcasts
• Any video (speech-to-text, beyond YouTube captions)
• TikTok
• Notion
• Google Docs
• EPUB

Until then, the general web pipeline may still extract something useful from some of these, with no guarantee of quality.