Two names for the same thing. Chroma and Milvus call the container a collection; Pinecone calls it an index. Either way it is the named space your embeddings live in and queries run against. An agent creates one before it adds vectors or searches.

What do I configure when creating one?

Usually the vector dimensions, the distance metric, and sometimes the index type (like HNSW) or an embedding model. Those choices shape how search behaves later, so getting them right at creation matters more than most setup steps, since changing them often means rebuilding the collection.

MCP servers that can create a vector collection

3 verified servers expose a tool that can create a collection in a vector database

A vector store organizes embeddings into collections, and before an agent can add or query vectors, the collection has to exist. Creating one sets up that container, with its dimensions and index, so the documents an agent indexes have somewhere to land.

These verified servers let an agent create a collection in a vector database.

Top pick

Chroma

Official

Chroma's official MCP server: manage collections and run semantic, metadata, and full-text search over a Chroma vector database.

vector-search

Tool:

chroma_create_collection

Chroma's chroma_create_collection makes a collection with optional HNSW configuration, a quick setup for a local or lightweight vector store.

Pick 2

Milvus

Zilliz

Official

Zilliz's official Milvus MCP server: vector, full-text, and hybrid search plus collection management over a Milvus vector database.

vector-search

Tool:

milvus_create_collection

On Milvus, milvus_create_collection creates one with a custom schema, indexes, and vector fields, for teams running it at scale.

Pick 3

Pinecone

Official

Pinecone's official developer MCP server: search indexes, manage records, rerank results, and look up Pinecone docs from your agent.

vector-search

Tool:

create-index-for-model

Pinecone calls it an index: create-index-for-model sets up a container wired to an integrated embedding model, so an agent can index text without embedding it first.

What to know

Creating a collection is the setup step before any retrieval-augmented work: the container holds the vectors, and its configuration decides how search behaves. What you set up varies, Chroma creates a collection with optional HNSW settings, Milvus one with a custom schema and index, Pinecone an index wired to an integrated embedding model. Pinecone calls its container an index rather than a collection, but the role is the same: the named space your embeddings live in and queries run against. Getting the dimensions and metric right here is what makes later search accurate.

A collection is infrastructure an agent provisions once, not per session, so remembering it matters. An agent that recreates a collection it already made either errors or, worse, starts a second empty one and indexes into the wrong place. Holding which collections exist keeps the agent adding to the right container instead of fragmenting its data.

Questions

Collection or index?: Two names for the same thing. Chroma and Milvus call the container a collection; Pinecone calls it an index. Either way it is the named space your embeddings live in and queries run against. An agent creates one before it adds vectors or searches.
What do I configure when creating one?: Usually the vector dimensions, the distance metric, and sometimes the index type (like HNSW) or an embedding model. Those choices shape how search behaves later, so getting them right at creation matters more than most setup steps, since changing them often means rebuilding the collection.