everypage-mcp
MCP server for EveryPage. Lets an AI assistant turn a PDF into a tracked share link, and tell you who actually read it — how far they got, how long they spent, and which page they stopped on.
"Share this proposal with Jane and Bob, then tell me tomorrow whether either of them read past the pricing page."
§ 01 Install
No install: the hosted server. EveryPage also runs this server at
https://everypage.co/mcp (Streamable HTTP, OAuth sign-in, no API key). In
Claude, open Customize → Connectors → Add custom connector and paste that URL;
in Claude Code, claude mcp add --transport http everypage https://everypage.co/mcp.
See everypage.co/integrations/claude.
Uploads there take an https link rather than a local file path.
This package (stdio). Get an API key from everypage.co/account — it
looks like ep_live_…. Then point your client at the server.
Claude Code
claude mcp add everypage --env EVERYPAGE_API_KEY=ep_live_... -- npx -y everypage-mcp
Claude Desktop, Cursor, VS Code, and anything else that reads a JSON config
{
"mcpServers": {
"everypage": {
"command": "npx",
"args": ["-y", "everypage-mcp"],
"env": { "EVERYPAGE_API_KEY": "ep_live_..." }
}
}
}
Requires Node 20 or newer.
§ 02 What it can do
Two prompts ship as slash commands: readership_digest (what happened over the last N days) and send_tracked_copies (one link per recipient, done right).
§ 03 Things worth knowing
Any way of naming a document works. Every tool that takes a file accepts a
UUID, a short link id, or a share URL pasted straight out of the browser —
including /embed/ links and ones carrying ?via= — so there is no need to look
up an id first.
Uploads stream. A local path is read off disk rather than into memory, so a 2 GB document on the Pro plan uploads without the process growing. Files are checked for the PDF magic bytes and against your plan's size cap before anything is sent.
A URL is fetched without your credentials. The url lane downloads from the
host you named using no EveryPage headers at all, then uploads the bytes. Your
API key never travels to a third-party host.
Readership is summarized, not dumped. The tool computes the finding — the
steepest drop in the reach funnel, the most re-read page — and puts the raw rows
in structuredContent for anything that wants to chart them.
Event cursors are per-stream. everypage_list_events returns a nextCursor,
and it is only valid for the same type. The three streams have independent id
sequences; crossing them silently returns the wrong window.
Deleting is reversible by default. everypage_delete_file trashes, and says
so; purge: true is the irreversible one. Tools that cannot be undone are marked
with MCP's destructiveHint, so clients that support confirmation prompts will
ask before running them.
§ 04 Configuration
§ 05 Plans
Some features are plan-gated, and a refusal will say which plan is needed.
everypage_account reports the current plan and what it locks.
The API allows 120 requests per minute per key.
§ 06 Not included
/files/claim and the Canva publishing lane (app plumbing, not agent work),
branding assets and hotspots (large Pro payloads with no natural agent use), and
the authenticated QR fetch — the qr_url every tool returns is public and needs
no credentials. Open an issue if you want any of them.
§ 07 Development
npm ci
npm run typecheck
npm test
npm run build
npm run inspector # drive the tools by hand in the MCP inspector
The tool surface is snapshotted in test/schemas.test.ts: renaming a tool or one
of its inputs fails CI on purpose, because it breaks every saved prompt and agent
config that names it.
Field names and output shapes are pinned by the cross-integration contract shared with the EveryPage Zapier, Make and n8n integrations — change that first.