Best Databases MCP Servers for Vector and RAG Stacks: A 2026 Comparison

The MCPFind databases category covers 317 servers as of May 2026. The vector and RAG subset is smaller than the Postgres slice, and one thing to call out up front: dedicated vector store servers for Qdrant, Weaviate, Pinecone, and Chroma haven't yet been verified to the MCPFind HEALTHY tier criteria. Those products run in production at plenty of shops; their MCP server implementations just haven't enriched into our HEALTHY classification yet. We'll revisit this list when their entries do.

What the directory does index today is a set of HEALTHY servers that hit the RAG data layer from very different angles. Airweave Search (6,092 stars) ships a full open-source retrieval pipeline backed by Vespa. Supabase MCP (2,556 stars) puts pgvector behind a single connection. ArcadeDB (762 stars) bundles a multi-model engine with a built-in vector embedding model. FalkorDB QueryWeaver (786 stars) takes graph-schema retrieval into natural-language database queries. Apple RAG MCP (112 stars) is a working hybrid search implementation served through MCP. This guide compares them on RAG data layer fit, not orchestration (LangChain, LlamaIndex, and friends are out of scope). See all servers in the MCPFind databases category.

Selection Criteria

Two filters shaped this list. First, HEALTHY classification in the MCPFind quality audit: 50+ GitHub stars (or vendor-official status) and a commit in the prior 90 days. Second, the server had to contribute to a vector search or retrieval pipeline in a concrete way. Anything that mentioned vectors only in passing, without explicit vector tool coverage, was excluded. Star counts come from the MCPFind directory snapshot on May 20, 2026. The comparison table uses a modified column set from the Postgres listicle: Hybrid Search replaces Read/Write, and the Best for column is specific to the vector use case.

Comparison Table

Airweave Search

Airweave calls itself an open-source context retrieval layer for AI agents and RAG systems. That description holds up. It sits between your data sources (50+ integrations including GitHub, Confluence, Notion, Airtable, Slack, and Dropbox) and your agent, handling ingestion, chunking, syncing, and retrieval so you're not rebuilding that pipeline for every new integration. Under the hood: Vespa for vector storage, PostgreSQL for metadata, Temporal for orchestration. The cloud version lives at app.airweave.ai. Self-hosted deploys through Docker Compose with a single ./start.sh. At 6,092 GitHub stars it's the highest-starred server in this guide, and one of the top HEALTHY entries across every category in MCPFind.

The MCP surface exposes retrieval, not storage management. Its primary tool is a natural-language search call against one or more Airweave collections, where a collection is a synchronized view of connected data sources, already chunked and indexed. An agent sends a query string and gets back the most semantically relevant chunks ranked by Vespa's vector similarity. Source filters narrow retrieval to specific integrations or document types, and the response can include full chunk text alongside source metadata. There's no tool for directly inserting raw vectors. Airweave manages indexing on the pipeline side.

Where this one earns its place is multi-source retrieval. If your agent needs to pull context from a mix of GitHub issues, Confluence pages, and a Slack archive in one query, Airweave handles the fan-out and unified ranking, no bespoke connector per source. The natural-language query model also keeps embedding generation off the client. Airweave embeds at ingest time, so the agent sends text and gets ranked chunks back. Teams already running LlamaIndex or LangChain as an orchestration layer can slot Airweave in as the retrieval backend those frameworks call through. It doesn't replace them.

Skip it if your retrieval surface is a single database or file store you already control. Airweave was built for the multi-source problem. One Postgres table with a vector column? Supabase MCP or ArcadeDB MCP gets you direct access without standing up a separate service. The hosted version also means Airweave runs embedding and sync timing on their infrastructure, which is a feature for most teams but a non-starter for anyone who needs fully air-gapped operation.

6,092 stars. Last push March 25, 2026. MIT license. Self-hosted option via Docker Compose.

Setup Snippet

Airweave MCP uses HTTP transport. The endpoint is the same whether you're on the cloud or self-hosted:

{
  "mcpServers": {
    "airweave": {
      "url": "https://mcp.airweave.ai",
      "env": {
        "AIRWEAVE_API_KEY": "your-api-key-here"
      }
    }
  }
}

For self-hosted deployments, swap the URL for your local instance address (default http://localhost:8080). See docs.airweave.ai for the current MCP endpoint and authentication setup.

Supabase MCP

Supabase runs on Postgres. Postgres with the pgvector extension is one of the most widely deployed vector stores in production. Supabase MCP (2,556 stars, github.com/supabase-community/supabase-mcp) wraps the Supabase management API alongside the database wire protocol, so an agent gets SQL execution plus Supabase platform features in one server. For vector workloads the flow is straightforward: enable pgvector through the Supabase extensions system, create a table with a vector column, then query similarity through execute_sql. All three steps happen inside a single MCP session. Supabase MCP gets full coverage of its broader database capabilities in the Postgres workflows listicle. This entry sticks to the vector use case.

The relevant tools for vector work are execute_sql, list_extensions, and apply_migration. list_extensions returns everything currently enabled on the project, pgvector included. execute_sql accepts any valid Postgres SQL, so a cosine similarity search through MCP looks identical to any other Postgres client: SELECT id, content, embedding <=> $1 AS distance FROM documents ORDER BY distance LIMIT 5. apply_migration covers the schema change needed to add a vector column to an existing table. One session confirms pgvector is on, adds the embedding column, runs a similarity query, and inspects the results without dropping into a SQL client.

If your team is already on Supabase as the primary database and doesn't want a second moving part for vectors, this is the obvious pick. A Postgres table with a vector column maintains cleanly, plays nice with row-level security, real-time subscriptions, and the rest of the Supabase stack you already operate. Teams starting a new RAG project and weighing pgvector against a dedicated vector store often land here for setup speed alone, no separate service to operate. Hybrid search works by combining vector similarity with tsvector full-text search in the same query, though that query is yours to construct.

The tradeoff against purpose-built vector stores is scale and indexing control. Past tens of millions of vectors, pgvector becomes harder to tune than a Qdrant or Weaviate, both of which expose explicit index configuration through their APIs. Corpus over a few million chunks, or hard latency requirements at scale? Plan the migration path before you start. For most early-to-mid-stage RAG stacks, Supabase on pgvector carries the load without a hiccup.

2,556 stars, last push March 24, 2026. Apache 2.0 license. Supabase free tier covers development workloads; production pgvector at scale runs on Pro or higher.

Setup Snippet

{
  "mcpServers": {
    "supabase": {
      "command": "npx",
      "args": ["-y", "@supabase/mcp-server-supabase@latest"],
      "env": {
        "SUPABASE_ACCESS_TOKEN": "your-personal-access-token"
      }
    }
  }
}

ArcadeDB MCP

ArcadeDB (github.com/ArcadeData/arcadedb) is a multi-model database that combines graph, document, key-value, time-series, search engine, and vector embedding models in one engine. The MCP server ships with ArcadeDB itself, no separate package install. Start ArcadeDB and the MCP surface is live. The vector model handles cosine, dot product, and Euclidean distance with configurable dimensions. The query layer accepts SQL, Cypher, Gremlin, GraphQL, and MongoDB query language. For teams that want vector retrieval next to graph traversals and document storage without operating three separate services, this is the only HEALTHY-tier server in this guide that bundles all three into one process.

The MCP tools follow ArcadeDB's command structure: query accepts SQL, Cypher, or Gremlin and returns JSON; command executes DDL or DML. A vector similarity search in ArcadeDB SQL starts by creating a type with a FLOAT array property and a vector index, then querying with the vectorNeighbors() function: SELECT @rid, content, vectorNeighbors('embeddings', $embedding, 5) AS score FROM Document. The graph traversal layer sits right next to that. If your documents have relationships encoded as graph edges, one query combines nearest-neighbor retrieval with path traversal. That combination isn't available in pgvector or in dedicated vector stores without external join logic.

The strongest case for ArcadeDB is the multi-model RAG pattern: documents with embeddings AND explicit relationships between them. Knowledge graphs where edges encode citation relationships. Code repositories where files have both vector embeddings and dependency edges. Product catalogs where items have embeddings and taxonomy links. In those scenarios, the graph + vector hybrid query capability cuts out join operations that would otherwise need multiple round trips. ArcadeDB runs from a Raspberry Pi to multi-server cloud deployments, per the project docs.

Pure vector-store workloads without relational or graph structure see less of a payoff here. ArcadeDB also takes more setup than a managed vector database (Java 17 or 21 runtime, self-hosted, no managed cloud tier as of this writing). Teams looking for zero-ops vector storage should look at Supabase MCP or Airweave instead. ArcadeDB is Apache 2.0, so commercial use is allowed without source disclosure.

~762 stars (MCPFind directory snapshot, May 2026). Last push March 24, 2026. Apache 2.0 license. Self-hosted only. See docs.arcadedb.com/#mcp-server for the current MCP configuration reference.

Setup Snippet

ArcadeDB's MCP server is built into the database. Point your client at the ArcadeDB HTTP endpoint:

{
  "mcpServers": {
    "arcadedb": {
      "url": "http://localhost:2480/<MCP_PATH>"
    }
  }
}

Replace localhost:2480 with your ArcadeDB host and port. The HTTP endpoint requires ArcadeDB 26.x or later. Check the ArcadeDB MCP docs for the current MCP endpoint path, auth configuration, and available query tools.

FalkorDB QueryWeaver

QueryWeaver (github.com/FalkorDB/QueryWeaver) converts plain-English questions into SQL by building a graph representation of your database schema first. That graph step is what differentiates it. Instead of handing raw DDL to a language model and hoping it maps table relationships right, QueryWeaver builds a graph of your schema (tables as nodes, foreign keys as edges) and uses that graph to resolve ambiguous query intent before generating SQL. Across normalized schemas with multiple joins, that approach tends to land more accurate SQL than approaches that ask the model to infer join paths from its training data. FalkorDB, which maintains QueryWeaver, is a graph database built on Redis.

QueryWeaver's MCP tools track the NL2SQL workflow end to end: list_databases enumerates connected databases, connect_database sets the active one, database_schema returns the schema graph, and query_database takes a natural-language question, generates SQL, executes it, and returns the result. For teams exploring RAG over structured data, where the "corpus" is a relational database rather than a document store, graph-schema-aware SQL generation is a meaningfully different retrieval pattern from embedding-based search. Retrieval here means finding the right rows, not the nearest vectors.

The right fit for QueryWeaver is when the retrieval problem is structured. "Which customers placed orders above $500 in Q1" is the use case, not "find documents related to customer churn." It also pulls weight when your team has tried using a language model for NL2SQL against a complex schema and watched join resolution be the failure mode. The graph-schema intermediate representation goes after that failure mode head-on. Teams already running FalkorDB as a graph store get an extension of that investment into the query authoring layer.

What QueryWeaver is not: a vector similarity tool. No embedding index, no nearest-neighbor search, no semantic retrieval in the vector sense. Teams building document-retrieval RAG pipelines should look at Airweave or Supabase pgvector instead. QueryWeaver is also AGPL-3.0, which carries copyleft obligations for distributed applications. If you're shipping a commercial product on top of it, check FalkorDB's commercial license terms before you build.

~786 stars (MCPFind directory snapshot, May 2026). Last push March 22, 2026. AGPL-3.0 license. Cloud at app.falkordb.cloud, self-hosted via Docker.

Setup Snippet

QueryWeaver exposes an HTTP MCP surface. The simplest way to run it locally is via Docker:

docker run -p 5000:5000 -it falkordb/queryweaver

Then add it to your MCP client config:

{
  "mcpServers": {
    "queryweaver": {
      "url": "http://localhost:5000/mcp"
    }
  }
}

For production or the managed cloud endpoint, refer to the QueryWeaver documentation for the current MCP URL and required environment variables.

Apple RAG MCP

Apple RAG MCP (github.com/BingoWon/apple-rag-mcp, slug com-apple-rag-mcp-server) indexes 370,000+ Apple Developer documentation pages and 1,300+ WWDC video transcripts, and exposes them through three search modes over MCP: semantic (vector similarity), keyword (BM25-style term matching), and hybrid (both combined with reranking). The corpus is Apple docs only, so this is a specialist tool, not a general-purpose RAG server. The reason it's in this guide is that it's the most complete production-grade RAG MCP implementation in the directory. The hybrid search pipeline, the reranking step, and the structured chunk response format are all patterns that transfer to any RAG stack you're designing.

Its tool surface is deliberately narrow. One search tool handles all three modes: pass a query string and a mode flag (semantic, keyword, or hybrid), get back ranked chunks with source URLs and relevance scores. Hybrid mode runs both retrieval paths and applies AI reranking to the merged result list. Response design favors token efficiency: no boilerplate metadata, no oversized context dumps, just the chunks the query needs. The server runs as a hosted HTTP endpoint at mcp.apple-rag.com. No API token needed for development-volume queries. An optional MCP Token (from apple-rag.com) lifts the rate limits.

The use case is tightly scoped: iOS and macOS developers on Claude or Cursor who want agents that know the latest SwiftUI APIs, deprecations, and framework changes without hallucinating off training data older than the current SDK. For that audience, this is the lowest-friction option in the directory. One config block, no account required, instant results. If your team is building a RAG server from scratch, the hybrid search implementation is also a useful reference for exposing all three retrieval modes through a single MCP tool cleanly.

Skip this when your RAG corpus is anything other than Apple Developer documentation. There's no ingestion API, no collection management, no support for custom data. General-purpose hybrid search is what Airweave is for.

~112 stars (MCPFind directory snapshot, May 2026). Last push March 20, 2026. License unconfirmed in the directory snapshot, check the GitHub repo before production use. Hosted endpoint only.

Setup Snippet

Apple RAG MCP uses HTTP transport via the hosted endpoint. No package install needed:

{
  "mcpServers": {
    "apple-rag-mcp": {
      "url": "https://mcp.apple-rag.com"
    }
  }
}

For higher rate limits, add your MCP Token from apple-rag.com to the request headers. Claude Desktop and Cursor both handle this through the headers field in the MCP config. Refer to your client's documentation for the exact config format.

How to Choose: Decision Framework

The first split is managed retrieval versus database-level vector storage. If your team doesn't want to operate a retrieval pipeline, Airweave runs ingestion, chunking, syncing, and search as a managed layer. Connect your data sources, point the MCP server at your Airweave instance, and retrieval is reachable through a single natural-language query tool. If you're already operating a database and want to bolt vector capabilities on, Supabase MCP (pgvector) and ArcadeDB MCP (built-in vector model) give you retrieval primitives without a separate service. The pick between those two usually comes down to whether you're already on Postgres (Supabase MCP, covered in depth in the Postgres workflows listicle) or need multi-model capabilities alongside vectors (ArcadeDB).

FalkorDB QueryWeaver covers a structurally different retrieval problem: natural-language queries over structured relational data with complex joins. RAG corpus that's a normalized database rather than a document store? Past NL2SQL attempts that fell over on join-path resolution? QueryWeaver's graph-schema approach is worth a look. It's not a replacement for the other servers in this list. It covers a different retrieval pattern.

For iOS and macOS teams, Apple RAG MCP fixes one persistent problem: agents hallucinating about deprecated APIs or missing framework changes. Its hybrid search implementation (semantic plus keyword plus reranking) is a useful reference for the retrieval pattern production RAG stacks lean on. Studying how it structures its tool surface pays off whether or not you deploy the server itself.

One coverage gap worth naming directly: dedicated vector store MCP servers for Qdrant, Weaviate, Pinecone, Milvus, and Chroma aren't yet in the MCPFind HEALTHY tier as of May 2026. These stores are heavily used in production. Their MCP server implementations just haven't been verified to our HEALTHY tier criteria yet. We'll revisit when those entries enrich. Community implementations on GitHub get covered from an ML workflow angle in the MCPFind ai-ml category post. For a purpose-built vector store MCP server with collection management, index configuration, and bulk upsert tools, that post covers the current state of those options.

Browse all 317 servers in the MCPFind databases category with the star filter set to 40+ to see the current adoption tier. The vector and RAG slice is growing as more purpose-built servers reach production quality.

Related Servers in This Category

Next Steps

• Best databases MCP servers for Postgres workflows
• Best MCP servers for AI and machine learning workflows
• Browse all databases MCP servers on MCPFind
• What is MCP and how does it work?