MCP server vs vector database is not an either or choice, because they do different jobs. A vector database stores embeddings and finds the closest matches to a query; it is where your knowledge lives. An MCP server is a small program that hands an AI agent a list of tools it can call, using the Model Context Protocol. The two meet when you put a vector database behind an MCP server, so an agent such as Claude Code or Cursor can search and save memories on its own. You need the database whenever you need semantic search. You need the MCP server only when an agent, not your own code, should decide when to search.

This guide explains each piece, shows how they stack, lists the MCP servers the main vector databases publish, and covers when to skip MCP and query the database straight from your code. We checked every server against its official repository or docs on October 31, 2026, and ran the Qdrant and Chroma servers ourselves. If MCP itself is new to you, start with what is MCP.

MCP server vs vector database at a glance

ItemVector databaseMCP server
What it isA database for embeddings plus metadataA program that exposes tools, resources and prompts to an AI client
Main jobStore vectors and run similarity searchLet a model discover and call actions
Who calls itYour code, through a client library or HTTP APIAn MCP host such as Claude Code, Cursor or VS Code
Holds data?Yes, that is the pointUsually not; it wraps something else
ExamplesQdrant, Chroma, Pinecone, Weaviate, Milvus, pgvectormcp-server-qdrant, chroma-mcp, Pinecone MCP, a server you write
Replaces the other?NoNo

The short version: the vector database is the storage and search engine, and the MCP server is one possible doorway into it.

What a vector database does

A vector database stores items as lists of numbers (embeddings) produced by an embedding model, next to the original text and some metadata. When a query comes in, it is embedded the same way, and the database returns the stored items whose vectors are closest. That is how retrieval augmented generation (RAG) finds the paragraphs to put in a prompt, and how agent memory finds the note a user wrote last week.

Good vector databases also filter on metadata (only this user, only this project), combine keyword scoring with vector scoring (hybrid search), and isolate tenants. Our best open source vector database roundup compares six of them, and Weaviate vs pgvector shows how much filtering behavior can differ between two engines.

What an MCP server does

The Model Context Protocol specification defines how an AI application (the host) talks to servers that offer three kinds of things: tools the model can call, resources it can read, and prompts it can reuse. A server can wrap anything: GitHub, a file system, a browser, or a database. The host lists the tools, shows their names and descriptions to the model, and runs a call when the model asks for one.

So an MCP server has no search engine of its own. It is an adapter with a fixed menu. For the difference between that menu and an ordinary HTTP API, see MCP server vs API.

How they work together

The common pattern is a vector database behind an MCP server:

  1. You run the database (a local folder, a Docker container or a cloud cluster).
  2. You start that database's MCP server and point it at the database with environment variables or flags.
  3. You register the server in your agent's MCP settings.
  4. The agent now sees tools such as "store" and "find" and calls them when it judges they help.

This is the quickest way to give a coding agent long term memory or let a chat client search your documents. The tradeoff is control. In a RAG pipeline your code decides when to retrieve, how many results to fetch and which filter to apply. Behind MCP, the model makes those calls, limited to whatever arguments the server exposes.

Vector database MCP servers you can use today

Most major vector databases now publish an official MCP server, but they differ a lot in what they let an agent do.

Qdrant. mcp-server-qdrant (0.8.1 on PyPI) exposes two tools, qdrant-store and qdrant-find. It works against a Qdrant server or a local folder, embeds text with FastEmbed (the only provider it supports so far, defaulting to all-MiniLM-L6-v2), returns up to 10 results by default, and drops the store tool when you set QDRANT_READ_ONLY=true. Our Qdrant guide shows how to add it to Claude Code.

Chroma. chroma-mcp (0.2.6) is a full management surface: list, create, modify and delete collections, then add, query, get, update and delete documents. It connects to an in memory, persistent, self hosted or Chroma Cloud client, and its query tool accepts where metadata filters and where_document text filters.

Pinecone. The Pinecone Developer MCP server (@pinecone-database/mcp 0.4.0, released October 29, 2026) needs Node.js 22 or later and a Pinecone API key for anything beyond searching Pinecone's docs. Its README says it is meant for coding assistants: it can create indexes, upsert and search records, run a search across several indexes and rerank. It does not support search with your own vectors, so it works with indexes that embed text for you. Pinecone points to a separate Assistant MCP for feeding knowledge base context to an assistant.

Weaviate. Weaviate builds its MCP server into the database at /v1/mcp on the REST port. It has four tools: read a collection's config, list tenants, run a hybrid query (with tenant and filter arguments) and upsert objects. Self hosted, it is off by default and read only until you set MCP_SERVER_WRITE_ACCESS_ENABLED=true. On Weaviate Cloud it is always on and the upsert tool is available until you turn on the read only switch. More in what is Weaviate.

Milvus. Zilliz maintains mcp-server-milvus with full text, text similarity, vector and hybrid search tools, a filter query tool, and tools to create collections, insert data and delete entities. Note that its milvus_vector_search tool takes a raw query vector as an argument, so the client has to produce the embedding first; the text search tools are the ones a chat agent can call with plain words.

pgvector. There is no pgvector specific server. You reach it through a Postgres MCP server. The original reference Postgres server now sits in the archived reference servers repository and its npm package is marked as no longer supported. Postgres MCP Pro is a maintained option with a restricted mode that limits the agent to read only transactions, and an unrestricted mode for development.

What we saw when we ran two of them

We drove both servers from a small script using the official MCP Python SDK, the same way an agent host would.

With mcp-server-qdrant, we stored a note about a staging server (tagged project alpha) and a note about a lunch order (tagged project beta), then asked "how do I reach staging". The staging note came back first, which is right, but the lunch note came back too, because the find tool takes only a query argument and returns up to its limit. With this setup the agent cannot ask for project alpha only, and metadata is returned for the agent to read, not used as a filter. If you need per user or per project scoping, give each scope its own collection or server, or write your own server with the filter built in; our MCP server in Python guide shows how.

With chroma-mcp, the same two documents and a query with n_results of 1 returned only the staging note. The server listed 13 tools, including one to fork a collection and others to delete collections and documents, and none carried MCP annotations such as a read only or destructive hint. One install note: a fresh uvx chroma-mcp@0.2.6 failed at startup with a pydantic import error on the day we tested. Adding --with "pydantic<2.12" to the uvx command fixed it.

When to query the vector database directly instead

MCP is the wrong layer when retrieval is a fixed step of your product rather than an agent's choice. Skip it and call the database from your code when:

  • Every request needs retrieval. A support bot that always looks up the docs should retrieve before the model runs. Making the model decide adds a round trip and a chance it skips the search. Our RAG pipeline with LangGraph tutorial builds this path.
  • Filters are a security boundary. Tenant or user filters should come from your session data, never from arguments the model fills in.
  • You need tuned search. Hybrid weights, rerankers, score thresholds and pagination are easier to control in code than through a fixed tool schema.
  • You are serving many users. An MCP server launched by one person's editor is a developer tool, not a shared retrieval service.

MCP fits best when a person is working with an agent: a coding agent saving project notes, a chat client searching a team's docs, or a developer asking an assistant to inspect an index. For deciding between a plain vector store and a full memory layer, read our AI agent memory comparison.

Safety checklist for vector database MCP servers

Treat a database MCP server like any tool with write access, because several of them can delete data:

  • Start read only. Use QDRANT_READ_ONLY=true, leave Weaviate write access off or connect with a viewer key, and run Postgres MCP Pro in restricted mode.
  • Use a scoped key. Give the server its own API key with the narrowest role the database offers, not an admin key.
  • Watch for schema changes. Weaviate's docs warn that with auto schema on, an upsert to a mistyped collection name creates a new collection instead of failing.
  • Pin versions. Run a fixed version (uvx mcp-server-qdrant@0.8.1) so a new release does not change your tools overnight, and pin dependencies when a fresh install breaks, as chroma-mcp did for us.
  • Keep it local first. Our guide to running an MCP server locally covers the MCP Inspector and transport options for testing before you connect a real agent.

Which should you use?

Use a vector database whenever you need to search by meaning; that part is not optional for RAG or agent memory. Add an MCP server on top when you want an agent in Claude Code, Cursor, VS Code or a chat app to decide for itself when to search or save. Write your own small server when the official one lacks the filter or scope you need. And skip MCP for production retrieval where your code should control every query.

FAQ

Is an MCP server a database?

No. An MCP server is a program that exposes tools, resources and prompts to an AI client over the Model Context Protocol. Some servers wrap a database, but the server itself usually stores nothing; the data stays in Qdrant, Chroma, Postgres or whichever system it connects to.

Can an MCP server replace RAG?

Not really. An MCP server can expose retrieval as a tool, which turns a fixed RAG step into a search the agent chooses to run. The retrieval still happens in a vector database or search engine. For always on grounding, a RAG step in your own code is more predictable.

Which vector databases have an official MCP server?

As of October 31, 2026: Qdrant (mcp-server-qdrant), Chroma (chroma-mcp), Pinecone (the Developer MCP server and the Assistant MCP), Weaviate (built into the database) and Milvus (mcp-server-milvus from Zilliz). pgvector is reached through a general Postgres MCP server.

Do I need a vector database to build an MCP server?

No. Most MCP servers wrap an API, files or a command line tool and never touch embeddings. You need a vector database only if your server has to search by meaning, for example to recall past notes or find relevant documents.

Is the Qdrant MCP server enough for agent memory?

For one person and one project, yes: it gives an agent store and find tools against a local folder. In our test the find tool took only a query, so it could not filter by project. Use separate collections per scope, or a custom server, when isolation matters.