Marqo AI (Vector Search MCP Server) MCP. Manage your entire vector knowledge base via chat.

You're running an AI client—Claude, Cursor, whatever—and you need it talking to your internal data. This server lets your agent handle complex vector search and indexing for your knowledge base without you having to write any boilerplate API code. You manage everything from creating the index structure all the way through querying documents, just by talking to it.

The tensor_search tool runs natural language queries against a specified vector index. When you use this, your agent doesn't look for keywords; it finds meaning. It executes dense semantic similarity searches, so if you ask something conceptually related to a document but don't use the exact words, the system pulls up the right stuff.

This means your AI client can answer questions based on context, not just matching text blocks.

When you need to build or update your data source, you've got create_index. You use this to build an explicitly bounded vector index from scratch. It lets you set custom model settings and dimension constraints right out of the gate, giving you fine-grained control over your search architecture before a single document gets loaded.

Think of it like laying down the perfect foundation for your data warehouse.

Once the container is built, you fill it using add_documents. This tool takes structured JSON records and writes them directly into that existing Marqo vector index. The moment those documents are added, they're immediately available for semantic searching by your agent. You can keep feeding fresh information to your knowledge base through conversation.

To make sure your data stays tight, you use delete_documents. This tool lets you remove specific records from an index by targeting them with their unique IDs. If a document gets outdated or is no longer relevant, you don't have to re-index the whole thing; you just delete it precisely.

Keeping track of what you've got is critical. The list_indexes tool shows every single vector index that currently exists on your Marqo instance. You run this when you want a full map—a quick inventory—of all your data collections before you try to query anything. It lets you know exactly what boundaries you're working within.

When you need an operational checkup, get_index_stats pulls detailed configuration metrics for any given index. You can see the document count, what embedding model it used, and its overall schema mapping. This is how you audit the health of your data source without having to manually count things or dive into a dashboard.

Essentially, you use this server so your agent talks directly to Marqo's core functionality: you define the index structure with create_index, populate it with fresh JSON using add_documents, run meaning-based searches with tensor_search, and then maintain that data—getting stats with get_index_stats, deleting garbage with delete_documents, and keeping tabs on everything with list_indexes.