My working routine lately looks like this: I open Claude and start thinking out loud. I describe a concept, turn an idea over, ask for an analysis or an article draft — and Claude, meanwhile, walks our knowledge base by itself, directly through MCP. In the previous article (Graphs in Meldoc) I talked about graphs and semantic connections between documents. Today — about how an agent uses those graphs. And why it’s fast.

Don’t drag the whole file

The usual picture: an agent gets access to the docs — and drags in the entire document. The whole file into context, the whole file back. We do it differently, and this is fundamental.

Our MCP supports partial reads and partial updates: the document isn’t hauled back and forth. Search runs over those same graphs from the previous article, and the result is structured smartly — the agent receives not a document but a small chunk. And the chunk size is decided by the LLM itself: needs more context — it increases it; a couple of paragraphs is enough — it shrinks it.

Search: by meaning and by letters

Let me say separately how this search works — because here we deliberately walked away from the classics. A classic RAG pipeline works like this: you pre-slice all documents into pieces, compute embeddings, store them in a separate store — and the model gets pieces of that snapshot in response to a question. I learned from my own experience: the snapshot has to be constantly rebuilt, that costs money, and between rebuilds it quietly goes stale.

We went another way: hybrid search over the live base. Vector search — by meaning — works in tandem with full-text search — by exact words. Ask “how to restrict access” — semantics finds the document about tokens even if the word “restrict” isn’t there. And when you’re looking for a specific name like parentAlias, exact search won’t let you miss. Each covers the other’s weaknesses.

And the main thing: no separate pipeline, no rebuilds. Whatever was fixed in the docs is what gets searched, immediately. The agent receives a fresh chunk on request, not a retelling from yesterday’s snapshot.

Next — my favorite moment. Having found the right place, the agent pulls in similar and related documents with one extra call — “where else is worth looking.” One step — and the context is topped up, without wandering around the base.

And when the agent reads a document, it arrives together with the list of glossary terms used in it. That’s how the LLM learns our slang and naming — the one accepted in our project.

Writing is surgical too

MCP works in two modes: read-only — or read and write. In the second case, editing can be partial: by blocks and whole sections, without rewriting the document. To change one sentence is to change one sentence.

How we live with it ourselves

The best proof is right in front of me: this series of articles is assembled by a team of AI agents, and all of their work with the knowledge base goes through that same MCP.

One agent gathers the material: searches the workspace and selectively reads the documents it needs, without dumping the whole base. Another verifies every claim about the product with live calls — and genuinely catches discrepancies: several phrasings in this series were fixed exactly that way. A third applies edits as pinpoint replacements: read the document, look at the preview, apply. Three steps, not “download all the text, rewrite it, upload it back.” And if someone changed the document in parallel — the edit simply won’t apply, nothing gets clobbered.

The humans haven’t gone anywhere: agents gather, verify and edit, but the final word is ours. Watching this kitchen from the inside, I can see: the “fast” in the title isn’t magic, it’s tool discipline. Partial operations plus preview make agent work with docs fast and safe at the same time.

And that’s not all

There’s a lot more of interest here. This toolkit already has ready-made applied scenarios: a documentation audit in one call (“We should clean up the docs”) and a ritual for documenting a legacy project (Don’t document legacy. Close the holes). And how MCP fits into the bigger picture — in Three copies of one truth.