Guides8 min read

Best MCP Servers for Data Scientists in 2026

Top MCP servers for data science workflows: Jupyter notebooks, SQL databases, BigQuery, Snowflake, Hugging Face, and more. AI-powered data analysis starts here.

By MyMCPTools Team·

Data science is drowning in context switching: a Jupyter notebook here, a SQL query there, a literature search somewhere else. MCP servers change that. Your AI assistant gets direct access to your notebooks, databases, and data pipelines — and suddenly exploratory analysis feels like a conversation.

Here are the MCP servers that actually move the needle for data scientists.

1. Jupyter MCP Server — Notebooks in Your AI Workflow

The Jupyter MCP server gives your AI assistant read and write access to running Jupyter kernels. You can ask your AI to inspect a DataFrame, fix a broken cell, or explain what a function does — all with real notebook context, not a pasted snippet.

Key capabilities:

  • Execute code in a live Jupyter kernel and retrieve outputs
  • Read and modify notebook cells programmatically
  • Inspect variables, DataFrames, and in-memory state
  • Restart kernels and manage environments

Best for: Anyone doing exploratory data analysis in Jupyter Lab or Notebook. This is the server that makes AI genuinely useful in a data science context.

2. PostgreSQL MCP Server — SQL You Can Actually Talk To

Most data scientists write SQL daily. The PostgreSQL MCP server gives your AI your actual schema — tables, columns, foreign keys, indexes — so it can write accurate queries instead of generic templates that break on your data model.

Key capabilities:

  • Full schema introspection (tables, columns, types, constraints)
  • Read-only query execution with safe defaults
  • Multi-schema and multi-database support
  • Explain plan analysis for query optimization

Best for: Data scientists working with production PostgreSQL databases or analytical replicas.

3. DuckDB MCP Server — Analytical SQL at Laptop Speed

DuckDB has become the go-to OLAP engine for local data science. The DuckDB MCP server lets your AI run analytical queries directly against Parquet files, CSVs, and in-memory datasets — no cloud warehouse needed.

Key capabilities:

  • Query Parquet, CSV, and JSON files directly without loading
  • In-process analytical queries with full SQL support
  • Window functions, CTEs, and advanced analytical SQL
  • Direct integration with pandas DataFrames

Best for: Data scientists who work with large flat files and want fast local analytics without spinning up a cloud warehouse.

4. BigQuery MCP Server — Petabyte-Scale Analysis via AI

Google BigQuery handles petabyte-scale analytics, but writing correct BigQuery SQL from memory is painful. The BigQuery MCP server gives your AI your actual datasets and schemas, enabling accurate query generation for even complex analytical workloads.

Key capabilities:

  • Dataset and table schema discovery across projects
  • Query execution with cost estimates before running
  • Support for BigQuery ML queries and functions
  • Partitioned table awareness for cost-efficient queries

Best for: Data scientists at companies using BigQuery as their cloud data warehouse.

5. Snowflake MCP Server — Enterprise Data Warehouse Access

Snowflake's multi-cluster architecture and share-based data marketplace make it the enterprise data warehouse of choice. The Snowflake MCP server brings your data model into context so your AI understands your schemas, stages, and warehouses.

Key capabilities:

  • Schema and table introspection across databases and schemas
  • Query execution with warehouse selection
  • Time travel query support for point-in-time analysis
  • Stage and file format awareness for ETL workflows

Best for: Data scientists working in enterprises running Snowflake as the central data platform.

6. Hugging Face MCP Server — ML Models and Datasets in Context

Hugging Face hosts over 500K models and 100K datasets. The Hugging Face MCP server lets your AI browse models, compare benchmarks, and pull dataset cards without leaving your workflow — critical when you're evaluating which pre-trained model to fine-tune.

Key capabilities:

  • Model search by task, architecture, and benchmark score
  • Dataset discovery with sample data previews
  • Model card and README access for documentation
  • Download links and code snippet generation

Best for: ML engineers and data scientists who work with pre-trained models and need to evaluate options quickly.

7. Databricks MCP Server — Unified Analytics Platform

Databricks spans data engineering, ML training, and analytical SQL in one platform. The Databricks MCP server gives your AI access to your Unity Catalog, notebooks, and SQL warehouses — enabling end-to-end data workflows without context switching.

Key capabilities:

  • Unity Catalog metadata browsing (tables, volumes, functions)
  • SQL warehouse query execution
  • Notebook cell inspection and editing
  • MLflow experiment and run access

Best for: Data teams using Databricks for the full stack from ingestion to ML training.

8. arXiv MCP Server — Research Literature at Your Fingertips

Data scientists rely on research papers for new techniques, architectures, and benchmarks. The arXiv MCP server lets your AI search and retrieve papers by keyword, author, or category — so you can discuss methodology with full paper context.

Key capabilities:

  • Full-text search across arXiv categories (cs.LG, stat.ML, etc.)
  • Abstract and metadata retrieval
  • Citation and reference lookup
  • Recent papers feed by category

Best for: Researchers and data scientists who need to stay current with ML literature or validate methodology against published work.

Recommended Data Science MCP Stacks

  • Local exploration: Jupyter + DuckDB + arXiv (notebook workflows + fast file analytics + research)
  • Cloud analytics: BigQuery or Snowflake + PostgreSQL + Hugging Face (warehouse + transactional + model discovery)
  • ML research: Jupyter + Hugging Face + arXiv + Databricks (full research-to-deployment pipeline)
  • Full data stack: All of the above — your AI has context across every layer from raw data to deployed model

Browse all Database MCP servers and AI & ML MCP servers on MyMCPTools. For related reading, see Best MCP Servers for Machine Learning Engineers and Best MCP Servers for Backend Developers.

Recommended Tools

Better Stack

Free Plan

Get alerted when your APIs, browser tests, payment pipelines, or MCP server dependencies go down. Used by 100K+ developers.

Start monitoring free →

1Password

14-day Free Trial

Store and inject API keys, payment credentials, tokens, and file access secrets into your MCP server configs. Trusted by 150K+ developers.

Try 1Password free →

🔧 MCP Servers Mentioned in This Article

💻

JupyterLab

Control JupyterLab notebooks from AI assistants. Execute cells, inspect variables, visualize outputs, and manage kernels programmatically.

Local
🗄️

PostgreSQL MCP Server

The PostgreSQL MCP server is an official Model Context Protocol server maintained by Anthropic that gives AI assistants read-only access to PostgreSQL databases. By connecting Claude Desktop, Cursor, or VS Code to a running Postgres instance, developers can ask natural-language questions about their data schema, run exploratory SQL queries, inspect table structures, list available schemas, and analyze query results — all without leaving their AI chat interface. The server operates in read-only mode by design, preventing any accidental data mutations, making it safe to connect against production databases for reporting, debugging, and data exploration workflows. Core tools include executing SELECT queries, listing tables and schemas, describing column types and constraints, and inspecting indexes. Setup requires a running PostgreSQL instance and a standard connection string in postgres:// format. Install via npx using the @modelcontextprotocol/server-postgres package, passing your database URI as an argument. Teams use it to power data analysis conversations, generate schema documentation automatically, debug production data anomalies by asking Claude to inspect table contents, and build ad-hoc reports through natural-language SQL generation. Works with any PostgreSQL 12+ instance including Amazon RDS, Supabase, Neon, and self-hosted deployments.

Local
🗄️

SQLite MCP Server

The SQLite MCP server is an official Anthropic reference implementation that gives AI assistants direct, conversational access to SQLite databases — the world's most widely deployed database engine. Through natural language, you can ask Claude or Cursor to run SELECT queries, insert and update rows, inspect table schemas, create new tables, and generate business intelligence reports without writing a single SQL statement manually. Common use cases include exploring local data files, prototyping application schemas, auditing CSV imports, running ad-hoc analytics on app databases, and letting AI agents manage lightweight structured storage during agentic workflows. The server exposes tools for query execution, schema introspection, and memo-style business insights that synthesize query results into readable summaries. It requires a path to an existing .db file as a startup argument. Install with: npx @modelcontextprotocol/server-sqlite /path/to/your-database.db. Works with Claude Desktop, Cursor, VS Code, and all MCP-compatible clients. For developers who want AI to reason directly over structured data stored locally, the SQLite MCP server is the fastest path from question to answer without leaving your AI chat interface.

Local
🗄️

BigQuery MCP Server

Google's official BigQuery MCP integration ships as part of the MCP Toolbox for Databases (googleapis/mcp-toolbox, 15,800+ stars, formerly published under the genai-toolbox repo name before Google renamed it), a single Go-based server binary that speaks the Model Context Protocol for over a dozen Google Cloud and third-party databases. Rather than a BigQuery-only package, you run the shared toolbox binary with a `--prebuilt=bigquery` flag to instantly load BigQuery-specific tools — schema/table discovery (`list_dataset_ids`, `list_table_ids`, `get_table_info`), running arbitrary SQL via `execute_sql`, and dry-run query validation for cost estimation before executing — over stdio or as an HTTP/SSE server. The quickest install is `npx -y @toolbox-sdk/server --prebuilt=bigquery --stdio` in your MCP client config; it also ships as a downloadable binary and Docker image for teams that prefer not to run via npx. Authentication uses standard Google Cloud credential chains (Application Default Credentials, service account keys, or Workload Identity) rather than embedding a project-specific key. Toolbox also underlies Google's official SDKs for Python, JS/TS, Go, and Java, so the same server config can back both ad hoc AI-assistant queries ("show me the schema for the events table and the row count for last week") and production agent tools built with LangChain, LlamaIndex, or ADK. For teams that want a fully managed remote option instead of self-hosting, Google Cloud also offers managed MCP servers for its databases including BigQuery.

Local
🗄️

Snowflake MCP Server

Snowflake now ships a first-party Snowflake-managed MCP server (Generally Available) that lets AI agents securely query Snowflake accounts over Streamable HTTP without deploying any local infrastructure — you configure it to expose Cortex Analyst, Cortex Search, and Cortex Agents as callable tools, plus custom tools and governed SQL execution, all through Snowflake's existing RBAC. It supports MCP revision 2025-11-25 and is documented under Snowflake AI & ML > Cortex Agents in the official docs. Before this hosted option shipped, Snowflake Labs published a community-maintained local server (Snowflake-Labs/mcp) covering Cortex Search/Analyst/Agents, object management, and SQL orchestration via a YAML service-configuration file and the Snowflake Python Connector for auth (username/password, key pair, OAuth, SSO, MFA) — that repo is now deprecated in favor of the managed server, though its docs remain useful for understanding the tool surface. For teams who want a self-hosted, read/write SQL-focused alternative instead of the managed offering, isaacwasserman/mcp-snowflake-server (community, 183+ stars) exposes read_query/write_query, schema-listing, and table-description tools via uvx, with an --allow-write flag gating destructive operations and a memo://insights resource that accumulates discovered data insights across a session.

Local
🤖

Hugging Face

Connect to Hugging Face Hub APIs - search spaces, papers, explore datasets and models.

Live
🗄️

Databricks

Connect to data, AI tools & agents, and the rest of the Databricks platform using turnkey managed MCP servers.

Local
🗄️

DuckDB Cloud

Serverless analytical database MCP for MotherDuck (cloud DuckDB). Run OLAP queries on large datasets, query Parquet and CSV files, and share data workspaces.

Local
🔍

arXiv MCP Server

The arXiv MCP Server bridges AI assistants to arXiv's research repository through the Model Context Protocol, letting Claude, Cursor, VS Code, and other MCP clients search, download, and read academic papers directly inside a conversation. The core workflow is search_papers → download_paper → read_paper: search_papers queries arXiv with boolean, category (cs.AI, cs.LG, cs.CL, cs.CV, stat.ML, quant-ph, and more), and date-range filters while automatically respecting arXiv's 3-second rate limit; download_paper fetches a paper by its arXiv ID (HTML first, PDF fallback) and stores it locally, returning content_length/next_start metadata so clients can safely page through very large papers; read_paper then returns the full text as markdown, with start/max_chars pagination for long documents. list_papers shows everything downloaded locally, and semantic_search searches across that local collection. The server also ships a "deep-paper-analysis" prompt that walks an assistant through executive summary, methodology, results, and future-research-direction analysis for a given paper ID. Install with `uv tool install arxiv-mcp-server` (NOT npm — an unrelated third-party package squats the same name on npm) or via the one-click Claude Desktop .mcpb bundle; a Streamable HTTP transport is available for server deployments. The README explicitly flags that arXiv paper content is untrusted external input and warns about prompt-injection risk (OWASP LLM01/AG01) when feeding raw paper text into agentic pipelines with tool access.

Local
🔍

Semantic Scholar MCP Server

The Semantic Scholar MCP Server is a FastMCP-based Python server that connects Claude, Cursor, VS Code, and other MCP clients to the Semantic Scholar API — the Allen Institute for AI's corpus of 200M+ academic papers — so an assistant can search literature, trace citation networks, and pull author profiles directly inside a conversation instead of guessing from training data. On the paper side it exposes full-text and title-based search with advanced filtering, multi-strategy ranked search, single- and multi-paper recommendations, and efficient batch detail retrieval with customizable field selection. Its citation-analysis tools walk the citation graph in both directions — citations and references — with citation context and influence signals, which is what makes it a strong fit for literature-review and citation-tracing agents rather than one-off lookups. Author tools cover author search, profile details, publication history, and batch author retrieval. It works unauthenticated for light use, but setting a SEMANTIC_SCHOLAR_API_KEY environment variable raises rate limits for heavier workflows; the server handles rate-limit compliance, connection pooling, and graceful shutdown internally. Install with `pip install semantic-scholar-fastmcp` or run it with no install via `uvx semantic-scholar-fastmcp`, then point your client's config at the uvx command. A Smithery one-click install for Claude Desktop and a companion Claude Code skills bundle (expand-references, trace-citations, paper-triage) are also available. Requires Python 3.10+.

Local

📚 More from the Blog