MKP from VkusVill: QAs themselves fill the shared memory of AI agents

⚡ The gist in 5 seconds

• VkusVill has unveiled MKP — an internal knowledge base platform that gives AI agents shared memory about testing.
• Inside there are two kinds of bases: evawiki with daily auto-updates and per-team markdown bases; search is hybrid RAG with an index in Qdrant.
• Access: web chat, REST API, and MCP for Cursor, Claude Code, and Codex; permissions are tied to a personal API key.

🔍 What happened

Oleg Malyshev, one of the leaders of the testing stack at TekhVill, broke down the internal MKP platform — Multi knowledge platform — in VkusVill's blog on Habr. It's not a product for sale: the thing works inside the company, for testers and analysts. The goal is to bring knowledge bases and smart document search together in one place.

The pain was simple. A QA spends two hours digging through code, logs, and Jaeger, writes to the dev chat, and finds out that the data mart takes availability from cache, while a mock of stock balances was intercepted by a live proxy with the wrong priority. The knowledge is found — and stays in someone's head and in a thread you won't find a week later. The AI agent has it worst: it has no memory between sessions, every run starts from scratch.

There are two types of bases in MKP. evawiki is one for everyone, a copy of EvaWiki, updated automatically every day, but read-only: documents can only be edited through a separate EvaWiki MCP. markdown bases — there can be as many as you like, one per team: testing architecture, test run results, found bugs and workarounds. A document can be created, updated, deleted, found by meaning, or fetched by number. There are three channels: a web interface with the sections "Chat", "Documents", "Runs", "Users", "Profile"; a REST API; and MCP — for Cursor, Claude Code, Codex, and anyone who speaks MCP. Permissions are the same for everyone: login with password in the interface, a personal API key for REST and MCP. No access to a base — the operation is rejected in any channel.

Under the hood is RAG, but not a naive one. The platform mixes hybrid dense/sparse retrieval — semantic search and keyword search at once. For QA this isn't a luxury: questions are full of identifiers like GetExists, is_pickup, or CMA-14699, and pure vector search loves to lose such strings. Then multi-query kicks in — the model itself rewrites the question in several ways. At this point the source fragment cuts off, so I don't know exactly how the sentence ended. The index lives in Qdrant.

🧠 Our take

The timing is no accident. Over the past year, MCP has become the de facto socket that agents in Cursor, Claude Code, and Codex plug tools into. Once the socket appeared, it turned out there wasn't much to plug in: corporate knowledge bases about the product exist, but the working knowledge of "wait five minutes while the availability cache warms up" is nowhere to be found. Our assessment: MKP isn't about AI — it's about the fact that context turned out to be more valuable than the model.

What it changes. For the market — a knowledge base turns from a wiki page into a service with an API and access control. For the team — working knowledge finally gets a home separate from product documentation, and that's the right split: EvaWiki answers "what should be", while MKP answers "how to actually verify it on the test stand". For the wallet, the cost isn't in the license but in discipline: someone has to write to the base while the knowledge is fresh, otherwise the RAG will obediently produce a beautiful answer based on garbage.

The winners are teams with a stream of similar tasks and a well-established TestOps — they have something to put in. The losers are those who expect the platform to conjure up memory by itself: an empty base searches perfectly, there's just nothing to search. External knowledge base vendors get an unpleasant competitor here — not a product, but an internal build you don't have to buy.

⚠️ What we don't know

There isn't a single number: how many teams have connected, how many documents are stored, how much faster task analysis has become, how often the agent answers off-target. The source is a corporate blog, i.e. self-presentation, and we have no way to verify it. Nothing is said about answer quality or retrieval errors, nor about token costs or Qdrant hosting. Permissions are described in words, without details: how keys are issued, what about auditing, who can see other teams' markdown bases. And keep in mind the text cuts off mid-description of multi-query.

Prediction: within six months, at least one major test management vendor will ship an MCP server for its knowledge base — by analogy with how everyone once got a REST API. Easy to check via the release notes of Allure TestOps and its competitors.

Source: Habr, VkusVill company blog