MCP server

What it is

MCP — the Model Context Protocol — is how an AI assistant talks to tools outside itself. DocuGrip ships a small MCP server: your assistant asks it what PDF tools exist, and it answers.

You do not have to know any of that to use it. Once it is connected, you just describe what you need in your own words and your assistant sends back the page that does it.

What your assistant can do with it

Three things: list every live tool, find the tools that match a task, and look up one tool by name.

The search takes plain language, not keywords. "Combine two contracts", "my scan has no searchable text", "get this under 2 MB" — all of those land on the right tool. The answer includes the URL in your own language, so a German user gets the German page.

Your file never reaches the server

This is the part worth reading twice. The server knows the catalogue and nothing else. It never receives a document, never opens one, and has nowhere to store one.

Every answer it gives is a link. The work happens in the browser tab that link opens, without ever reaching us — the same way every tool on this site already works.

An assistant wired to a server-side PDF service has to upload your document before anything can happen. Ours cannot, because there is nothing on our side to upload to.

Connecting it

The server speaks over standard input and output, one JSON message per line. Point your MCP client at it:

{ "command": "npx", "args": ["tsx", "src/mcp/server.ts"] }

There is no API key, no account and no sign-up. It makes no network calls of its own.

Built without an SDK, on purpose

The protocol surface we need is three methods: initialize, tools/list and tools/call. An SDK would have been larger than the part of it we use.

That is the same call this project made for Stripe, Supabase, CSV parsing and the service worker. A dependency you barely use is a cost, not a saving — and for you it means one less thing to install.