Drsti Precast: a private MCP server that lets AI assistants search company knowledge
How Gihonix built an authenticated MCP server for Drsti Precast, so AI assistants can list, search and read the company knowledge base, with a separate revocable key for each person.
The Gihonix team · · 3 min read
Key takeaways
- The server exposes three tools: list documents, ranked search and read a document in pages.
- Each person or client gets their own key, stored only as a hash and revocable on its own.
- New documents go live by adding a file to a folder, with no rebuild.
Drsti Precast, a precast construction company, wanted its product and technical knowledge to be usable by AI assistants. Gihonix built a private, authenticated MCP server that lets assistants such as Claude list, search and read the company knowledge base, with a separate revocable access key for every person.
The challenge
Company knowledge usually sits in documents: a product guide, technical notes, web pages. Staff search them by hand, and an AI assistant cannot reach them at all unless someone pastes the text into a chat. Pasting is slow, goes stale, and sends whole documents to places they were never meant to go.
The goal was narrower and safer: let an assistant ask the knowledge base a question and get passages back, while the company controls who can ask.
What we built
An MCP server. The Model Context Protocol (MCP) is an open standard that lets an AI assistant call a short list of named tools in another system. This server exposes exactly three:
- List documents. Every document with its id, title, type and size, with an optional filter.
- Search documents. Ranked search across passages using BM25, a standard text-ranking method, returning the matching text, the document id and the position.
- Read document. The full text of one document, delivered in pages so a long document does not flood the assistant.
The company's knowledge base, a PDF, was converted into searchable sections and loaded. HTML, Markdown and plain-text files are supported too.
How access is controlled
Every request needs a bearer key, or it is refused. There is one key per person or client, so one can be revoked without touching the rest, and a revoked key stops working at once. Keys are stored only as hashes, so reading the key file does not reveal them. Repeated failed attempts from one address are rate limited. Only a health check is open.
The server is ready to run behind HTTPS, and a Docker image is included, with the content folder and keys mounted so updates do not need a rebuild.
Keeping it current
New documents are added by dropping a file into a folder. The server rescans every couple of seconds, so a new document is searchable without a restart or a new build.
What changed
The team and its AI tools can get answers from the company's own documents, and the company decides who has access and can withdraw it per person. We are not publishing measured figures for this project.
Technology
Node.js, the Model Context Protocol, BM25 search, hashed bearer keys, and Docker.
Work with Gihonix
Gihonix is a software, AI and web development company based in Coimbatore, India. If you want AI assistants to use your own documents safely, see our MCP development service, read what an MCP server is, or contact us.