@page2ai/mcp — Web to Markdown for LLM Context
[npm version]
[CI]
[MCP Registry]
[license MIT]
No downloads badge on purpose: npm download counts are dominated by registry mirrors and
bots, and this README only shows numbers a human can trust.
One-click webpage-to-Markdown MCP server for ChatGPT, Claude, and Gemini.
Preserves code blocks with language hints, tables, and reading structure.
Fetch, extract, convert — one tool, one command, entirely on your machine.
No API key, no account, no telemetry; the only network requests are to the URLs you pass it.
Companion to the Page2AI Chrome & Firefox extensions,
which convert the live page in your own browser with their own pipeline. This server is built on
@page2ai/core.
Quick start
This starts the stdio MCP server — it waits silently for a client, so wire it into one of
the clients below rather than running it bare. Requires Node.js 20+. No API key. If
npx ever serves you a stale cached version, run npx -y page2ai-mcp@latest once.
Updating. npx -y page2ai-mcp asks npm for the newest version each time the server
starts, but an MCP client keeps the server running until you quit it: quitting and
reopening the client is the update. A version named in the config ([email protected])
stays until you change it. npx -y page2ai-mcp --version prints the version npx runs now.
page2ai-mcp is a thin unscoped wrapper around @page2ai/mcp,
released together with it; its dependency is a ^ range, which on 0.x versions admits
patch releases only.
Hosted endpoint (beta) — no install at all
For clients that only accept a remote MCP server URL (Streamable HTTP):
Same tool, same code (see remote/), rate-limited per IP.
Privacy trade-off: URLs sent to the hosted endpoint are fetched from that
server, so they transit infrastructure operated by the author (on Vercel).
Running locally via npx keeps everything on your machine — prefer it when you
can.
Claude Code
Or install the Claude Code plugin, which adds the same server plus a skill that tells the agent when to
use it and what to do with signed-in, blocked and very long pages (Claude Code 2.1.275 or later):
It asks for one optional, secret setting, the Agent Bridge key (see below). In claude -p runs allow the
tool by name: --allowedTools "mcp__plugin_page2ai_page2ai__page_to_markdown".
Claude Desktop
One-click bundle: download page2ai-mcp-<version>.mcpb from the
latest release, then in
Claude Desktop open Settings → Extensions → Advanced settings → Install Extension…
and pick the file.
Or add to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/,
Windows: %APPDATA%\Claude\), then fully quit and restart Claude Desktop:
Cursor
[Install MCP Server]
Or add to ~/.cursor/mcp.json (global) / .cursor/mcp.json (project):
VS Code (GitHub Copilot)
Or add to .vscode/mcp.json:
Codex CLI
Gemini CLI
Antigravity
Windsurf
Add to ~/.codeium/windsurf/mcp_config.json (macOS/Linux) or
%USERPROFILE%\.codeium\windsurf\mcp_config.json (Windows):
Zed
Add to settings.json (zed: open settings file):
Other MCP clients
Any client that speaks MCP over stdio works — Cline, Continue, JetBrains AI Assistant,
LM Studio and others. Point their MCP server settings at the same command:
npx -y page2ai-mcp.
Letting an AI agent install it
If an AI agent (Cline, Claude Code, Codex, …) is doing the setup for you, point it at
llms-install.md — a machine-oriented install guide with the exact
command per client and a verification step.
Tools
Long pages (0.7.0 and later)
A page longer than max_length characters (default 30,000, about 8,000 tokens) comes back in
parts. A part ends between blocks: never inside a code block or the frontmatter, and never right
after a heading. The first part ends with a note naming the start_index of the next part,
followed by the page's sections as start_index (length): heading, so an agent can read the one
section it needs:
Later parts come from the text the first part was cut from, kept in the server's memory for 10
minutes after the last part was read: the site gets no new request, Agent Bridge opens no second
tab, and offsets do not move. A page that fits in one part comes back exactly as before. Python's
built-in types page, about 286,000 characters in one result up to 0.6.3, now comes as a first part
of under 30,000 characters plus a list of its sections. The hosted endpoint returns whole pages.
When a page cannot be read (0.7.0 and later)
A fetch error keeps its first line and adds a second one with the next step. HTTP 401 and 403
(on 2026-10-30, npm, Stack Overflow, Medium and OpenAI refused the server's requests): pass the
page's HTML from a browser tool as html, or turn on Agent Bridge, since the same request will be
refused again; the step names PAGE2AI_BRIDGE_KEY when the server has no key, and says so when
another Page2AI server holds the bridge port. 404 and 410: check the URL. 429: wait. 5xx: try
later. A timeout: a larger timeout_ms. A PDF or another type that is not a page, the 10 MB cap
and a redirect loop name themselves, and a host name that does not resolve says so instead of
reading like an address-guard block. A page with no readable text (built by scripts in the
browser) and a fetch that landed on a sign-in page come back with a note giving the same steps.
Pages behind a sign-in or rendered by JavaScript
The server fetches without cookies and runs no page scripts. For a page that needs either:
-
Pass the HTML you already have (from Playwright, Chrome DevTools MCP or another browser
tool) as html, with the page's url: the HTML is converted, the URL is never fetched, and
relative links resolve against it. The frontmatter then carries provenance: "HTML supplied by the caller; the source URL was not fetched". An empty html is an error, not a fetch of url.
-
Or turn on Agent Bridge in the Page2AI extension (1.6.0 or later): popup → Agent Bridge
switch → allow access to websites. While the extension is connected, url pages open in a
background tab of your own browser, with its cookies and scripts, and come back as the
extension's Markdown (the call's include_images and include_frontmatter are honoured
there, not the popup's). The extension refuses sign-in providers' pages and anything that is
not http(s); a page it refuses is fetched by the server as usual, with a second text block
saying why; when the browser ends on another page (a sign-in redirect), a second text block
names the page it got. Private, loopback and cloud metadata addresses are refused before the
browser opens them, as by the server's own fetch, and so is a page the browser was redirected
to at one. The bridge listens on ws://127.0.0.1:3025 only.
The key (0.6.0 and later). The bridge works only with the key the extension shows: popup →
Agent Bridge on → Copy key. Give it to this server as PAGE2AI_BRIDGE_KEY: in Claude Desktop
(the .mcpb extension) paste it into Page2AI's setting Agent Bridge key; in a JSON config add
"env": { "PAGE2AI_BRIDGE_KEY": "<key>" } to the server entry; with Claude Code,
claude mcp add page2ai -e PAGE2AI_BRIDGE_KEY=<key> -- npx -y page2ai-mcp. Restart the MCP
client after. Several browsers or profiles: put their keys in the one variable, separated by
commas; the first to connect holds the bridge. Several MCP clients running Page2AI (Claude
Desktop and Cursor, say): the one started first holds port 3025 and the others run without the
bridge, so give the key to each of them. Without a key the server never sends a page to a
browser, and the popup and this server's log say what to paste where. On every connection both
sides prove they know the key (an HMAC exchange; the key never crosses the socket), the server
first: a program that took port 3025 while no server ran gets nothing from the extension, and a
program that connects to the server without the key cannot ask it for pages. Firefox needs no
extra setting since 0.6.0 (0.5.3 required the install's UUID in PAGE2AI_BRIDGE_ORIGINS).
New key in the popup, or reinstalling the extension, makes a new key: replace the old one with it.
Whoever reads your MCP client's config or the browser profile can read the key: the bridge is
no stronger than your user account.
Example prompts
1. Fetch documentation and answer questions:
"Use page_to_markdown to fetch https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking and summarize the three main use cases for extended thinking."
2. Extract API reference into a code snippet:
"Fetch https://ai.google.dev/gemini-api/docs/thinking with page_to_markdown, then generate a Python code sample using the thinking budget parameter."
3. Compare two documentation pages:
"Fetch both https://docs.anthropic.com/en/docs/prompt-engineering and https://platform.openai.com/docs/guides/prompt-engineering with page_to_markdown, then summarize the differences in approach."
What a call returns
Verbatim output of page_to_markdown(url="https://docs.anthropic.com/en/api/messages"),
captured 2026-10-18 with core 0.2.1, truncated:
YAML front matter (title, source, capture timestamp, page metadata, extractor provenance),
then the page as Markdown: headings, tables and fenced code preserved; navigation, header,
footer and sidebars dropped. Interface buttons are left out (core 0.2.1); up to 0.3.0 this page also printed «Copy page»
(with the glyph of its icon) and the «cURL» code-sample switcher, each twice. Up to 0.2.8 this URL returned the site's own
Markdown copy (extractor_source: "content-negotiation"); since 0.3.0 the server reads only
the page.
Configuration
No setup required — every argument has a sensible default. page_to_markdown accepts:
The hosted endpoint takes url only (the contract its ChatGPT plugin was reviewed with), caps
timeout_ms at 20000 and returns whole pages.
Environment variables of the local server, all optional:
Planned: a profile argument (auto | docs | marketing | research | dashboard | wordpress-marketing).
Privacy
@page2ai/mcp collects no data, sends no telemetry, and makes no external network calls
beyond the URLs you explicitly provide. Note that your AI client (Claude, ChatGPT, …)
still sends the extracted Markdown to its own model provider under that client's privacy
policy — that part is outside this server's control. See PRIVACY.md for details.
A long page's Markdown stays in the server's memory while its parts are read, until 10 minutes
after the last part was read or until the server stops; nothing is written to disk.
With Agent Bridge on, the extension sends this server the URL's Markdown as your browser
renders it, signed-in content included, over a WebSocket on 127.0.0.1; nothing else leaves
the machine, but the Markdown then goes to your AI client like any other result.
The optional hosted endpoint is different by nature: URLs you ask it to convert are
fetched from a server operated by the author (deployed on Vercel), and Vercel keeps its
standard operational logs. No accounts, no tracking. If that trade-off matters for your
use case, run the server locally — it is the same code.
Protocol revisions
Since 0.2.0 this server answers both MCP protocol revisions on the same stdio
connection:
serveStdio picks the era from the opening message and pins one server instance for
the life of the connection, so clients on older SDKs are unaffected. Note that a
message carrying no version claim is treated by the specification as a 2025-era
opening; server/discover without that _meta field is therefore answered with
-32601 rather than being upgraded.
Verify against a build with the raw JSON-RPC frames, without a client library:
Known advisories
None. Moving to the v2 SDK removed 90 transitive packages (117 production
dependencies down to 26) and with them the @hono/node-server advisory that
0.1.x carried through the monolithic @modelcontextprotocol/sdk; that package
pulled the HTTP and OAuth server stack even for a stdio-only server. npm audit
is clean as of 0.2.0.
Development
Test with MCP Inspector:
Support
Pick the channel by what you have, not by what is quickest to type:
About the author
Written and maintained by Igor Saevets — AI expert and founder of Page2AI. Full bio:
igorsaevets.github.io/page2ai-docs/about/.
These are identity and collaboration links, not the support queue. A bug reported in a DM is a bug
nobody else can find later, so anything you want fixed belongs in the table above.
Build provenance
An MCP server runs with the privileges of whatever launched it, so where the tarball came from
is a security question, not a formality. Releases from v0.1.2 onward are published from GitHub
Actions with npm provenance: each
version carries a Sigstore attestation naming the commit and workflow run that produced it,
recorded in the public Rekor transparency ledger and shown as a badge on npmjs.com.
Check it before you trust it:
License
MIT — see LICENSE. Copyright © 2026 Igor Saevets.
Related
