State of AI/ML tooling in MCP

The average AgentRank score across vector database tools is 38.4 — significantly below the index average of 29.7 (and much lower than database tools at 41.2). This isn't a quality problem with individual tools. It reflects recency: most AI/ML MCP servers launched in late 2025 or early 2026, so their freshness scores are good but contributor counts and dependent counts are still building.

Three patterns explain the landscape:

• Official vendor servers exist but are still early. Qdrant, Milvus, ChromaDB, and HuggingFace all have official MCP servers — none has hit critical mass yet.
• RAG tools score higher because they're general-purpose and attract more contributors. ForLoopCodes/contextplus (88.41) and agentset-ai/agentset (71.91) both have strong signals despite being newer.
• The freshness cliff matters here. chroma-core/chroma-mcp last committed September 2025 — almost six months ago. That's the biggest risk signal in this category.

Vector database MCP servers

The three major open-source vector databases — Qdrant, Milvus, and ChromaDB — all have official MCP servers. All three are authored by the database vendor. All three score in the 52–54 range.

Qdrant (52.68)

qdrant/mcp-server-qdrant is the official implementation from Qdrant. 1,273 stars — the most-starred vector DB server. Supports similarity search, payload filtering, and collection management. Last commit was January 2026, which is a mild freshness concern but not alarming. If you're running Qdrant, this is the only official path.

Milvus (53.59)

zilliztech/mcp-server-milvus from Zilliz covers both Milvus and Zilliz Cloud. Scored slightly higher than Qdrant's server despite fewer stars, due to better contributor signals. The last commit was December 2025, which is the most recent gap concern in this group. Still the right choice for Milvus deployments.

ChromaDB (52.05)

chroma-core/chroma-mcp is the official ChromaDB server. Last commit was September 2025 — the oldest of the three. That's a significant freshness penalty and is the reason it scores lowest despite being the official implementation. ChromaDB is popular for local development and prototyping; check the repo for recent activity before depending on it in production.

Bottom line for vector DBs: All three are official, all three score similarly. Pick based on which vector database you're already running. If you're starting fresh, Qdrant's server has the most stars and the most recent commit activity.

RAG and semantic search

RAG-focused tools score significantly higher than vector DB servers because they're used in more diverse workflows and attract more contributors. Two strong options:

ForLoopCodes/contextplus (88.41)

The highest-scoring tool in this category at 88.41. Context+ is built specifically for large codebase navigation: it combines RAG with Tree-sitter AST parsing and Spectral Clustering to build a hierarchical feature graph of your codebase. Not a general-purpose RAG tool — it's designed for agent workflows that need accurate code context. 1,459 stars, TypeScript, actively maintained with commits through March 2026.

agentset-ai/agentset (71.91)

The most general-purpose RAG platform in the index. 1,915 stars, supports 22+ file formats, built-in citations, deep research mode, and an MCP server interface. Open source with TypeScript. If you're building a knowledge base or document retrieval system for an agent, agentset is the strongest option that's not domain-specific.

paiml/paiml-mcp-agent-toolkit (88.57)

Written in Rust — unusual in the MCP ecosystem — and designed to make AI-generated code more deterministic. Focuses on code quality signals (complexity, dead code, dependency analysis) rather than retrieval. Scores 88.57 due to strong freshness and issue health. Particularly useful if your agent is generating or modifying code and you need quality checks in the loop.

Model access and inference

Direct model access via MCP is less developed than vector search or RAG. The only official vendor server with meaningful quality signals is HuggingFace.

huggingface/hf-mcp-server is the official HuggingFace server at 76.39. Covers model search, dataset access, and inference via the HuggingFace API. 207 stars, 10 contributors, TypeScript. The last commit was February 2026 — still fresh enough to depend on.

If you need to call models directly from an agent workflow, the HuggingFace server is the highest-quality option in the index. There are no official servers for OpenAI, Anthropic, or Mistral that score above 60 — those integrations are typically handled at the client layer rather than through MCP.

Recommended picks by workflow

Building a RAG pipeline

Start with agentset-ai/agentset for general document ingestion and retrieval, or ForLoopCodes/contextplus if your primary use case is code navigation. Pair with qdrant/mcp-server-qdrant if you want control over the vector store layer.

Semantic search over a codebase

ForLoopCodes/contextplus is purpose-built for this. It's the best-maintained and highest-scoring codebase search tool in the index.

Accessing HuggingFace models and datasets

huggingface/hf-mcp-server is the only option here with credible quality signals. It's official and actively maintained.

Using ChromaDB locally

chroma-core/chroma-mcp is the right tool but monitor the repo for activity — the September 2025 last commit is worth watching. Check for updates before pinning a version in production.