Use cases ·
MCP server for company documents: how it works
An MCP server for company documents lets an AI assistant search your company's files and answer with links to the full document and page. The answers are only as good as the retrieval behind the server: on SECQUE, a public test on SEC filings, Hippocampus got 81% of the questions right and the next best system 58%.*
On this pageWhat is an MCP server for company documents?
- What is an MCP server for company documents?
- How does an MCP server for documents work?
- How accurate are answers through an MCP server?
- Should you use an MCP server or upload files?
- Should you build your own MCP server?
- Who can see what through an MCP server?
- How the Hippocampus MCP server works
- Try it with your own documents
What is an MCP server for company documents?
An MCP server for company documents is a service that makes a company's documents available to AI assistants through the Model Context Protocol, an open standard for connecting assistants to tools and data. The assistant calls the server to search, open a document and read a passage, and then writes its answer from what it read. Claude and ChatGPT both connect to MCP servers, so you set up the documents once and use them from either.
How does an MCP server for documents work?
An MCP server for documents works in five steps.
- You import your documents into the server.
- You add the server's address in your assistant and sign in.
- You ask a question in the assistant, in plain language.
- The assistant calls the server, which returns the passages that answer the question.
- The assistant writes the answer and links to the full document and the page.
How accurate are answers through an MCP server?
Answers through an MCP server are as accurate as the retrieval behind it, because the protocol only carries the question and the passages. Two public tests on company filings show the spread. On SECQUE, 565 expert questions on 45 SEC filings, Hippocampus got 81.2% right and Azure AI Search 58.0%.* On FinanceBench, Hippocampus got 92.2% and OpenAI File Search 52.3%. One model wrote the answers for each system, so the gap is down to which passages the server returned.
Should you use an MCP server or upload files?
Upload files when you have a few documents and one person asking, and use an MCP server when the documents number in the hundreds, change over time or are shared by a team. An uploaded file is a copy that stays as it was on the day you uploaded it, inside one chat or project. A server keeps one memory that every assistant and agent asks, and you update it in one place. We compare the two in Hippocampus vs OpenAI File Search.
Should you build your own MCP server?
Build your own MCP server if you want to learn the protocol or have a small, stable set of documents. A basic server is quick to build, and open-source examples exist. The protocol is the easy part. The hard part is what the server returns: the right passage among thousands of documents, only from documents the user may see, and a refusal when nothing fits. Use a ready server when answers have to be right and many people depend on them.
Who can see what through an MCP server?
Through an MCP server, each person should see only the documents their role allows. Check that the server signs each user in individually and applies document access to every search. A server that runs on one shared key gives every user the same view of every document. Ask where the documents are stored and whether they are used to train models.
How the Hippocampus MCP server works
Hippocampus runs an MCP server over streamable HTTP with OAuth sign-in, so each user signs in as themselves and access follows their role. An MCP client can search, open folders, documents and sections, read the exact text and import files. Hippocampus refuses a question the documents do not cover. MCP and the API are included on every plan. It works with Claude, ChatGPT and other MCP clients, and people who prefer it can ask in the web browser instead. More in the MCP documentation.
Try it with your own documents
Create a free account, import your documents and connect Claude or ChatGPT today. The Free plan includes the MCP server and the API, holds 2,000 pages and needs no card. The Starter plan holds 25,000 pages for EUR 30 a month. Larger teams get the Corporate plan, with custom pricing, single sign-on and the option to host it themselves.
Get started free or book a demo.
* GPT-5.4 mini answering. See Evidence.