Ranked roundup · Databases
Best MCP Server for Database Access
The strongest MCP server for databases right now comes from Google: MCP Toolbox, an open-source server covering roughly twenty engines — Postgres, MySQL, SQL Server, MongoDB, ClickHouse, Snowflake, and the whole Google Cloud line among them. It sits at 16,288 GitHub stars with commits pushed the day we checked. The rest of the landscape splits two ways: database vendors ship their own official servers (Supabase, Neon, and MongoDB all do, some with hosted OAuth endpoints), and community projects cover plain self-hosted databases — Bytebase's DBHub as a lightweight multi-engine gateway, Postgres MCP Pro for single-engine depth. Note what does not exist: an official server for "databases" as a category. Anthropic's reference Postgres and SQLite servers were archived in 2025 and are unmaintained. Everything ranked below was verified on September 1, 2026 against public repos and vendor docs — official status, push recency, stars, auth model, capability coverage. No paid placement, and no invented benchmarks.
Facts checked September 1, 2026 · ranked on verifiable criteria · placement is never for sale
Founder, Automation Squad ·
The ranking
- Official · Google
Teams running multiple database engines, or anything on Google Cloud.
Prebuilt generic tools (execute_sql, list_tables, schema exploration) for roughly twenty engines — PostgreSQL, MySQL, MariaDB, SQL Server, Oracle, MongoDB, Redis, Elasticsearch, CockroachDB, ClickHouse, Couchbase, Neo4j, Snowflake, Trino, plus BigQuery, AlloyDB, Cloud SQL, Spanner, and Firestore — and a custom-tools framework (tools.yaml) for predefined structured queries, with connection pooling and OpenTelemetry built in.
Self-hosted: single Go binary, Docker, or npx, with stdio and HTTP transports. A separate managed offering exists, but only for Google Cloud databases. · Connection credentials via environment variables or tools.yaml config; integrated Google Cloud IAM for cloud-hosted sources. · Apache-2.0 · 16,288★
- The repo is mid-rename from genai-toolbox to mcp-toolbox, so older docs links 404.
- Custom tools mean learning the tools.yaml configuration format — more setup than a single-DSN server.
- The managed hosted version only covers Google Cloud databases; everything else you run yourself.
- Community
One small, token-conscious server for the common SQL engines with sane safety defaults.
execute_sql and search_objects across PostgreSQL, MySQL, MariaDB, SQL Server, and SQLite through one interface, with read-only mode, row limiting, and query timeouts as guardrails; optional explain_sql, health_check, and custom parameterized tools.
Self-hosted via npm or Docker; stdio and HTTP transports; includes a built-in Workbench web UI. · Single DSN connection string with credentials embedded; SSH tunneling and SSL/TLS supported. · MIT · 3,439★
- Not an official server from any database vendor — it is maintained by Bytebase, a database DevOps company.
- Credentials live in the DSN string, so use a least-privilege database user.
- SQL engines only — no MongoDB, Redis, or other NoSQL coverage.
- Official · Supabase
Anyone whose Postgres already lives on Supabase — the cleanest auth model in this roundup.
SQL execution, table and migration management, and project configuration for Supabase-hosted Postgres, with a read_only=true flag to strip mutating tools and project_ref scoping to lock access to one project.
Hosted remote endpoint — no local install; project-scoped MCP URLs can be generated from the Supabase dashboard. · OAuth 2.1 browser login on the hosted endpoint (mcp.supabase.com); personal access token for CLI and self-hosted setups. · Apache-2.0 · 2,884★
- Supabase projects only — not a general-purpose Postgres server.
- The default tool set includes broad project-management powers; Supabase itself tells you to read its security best-practices guide before connecting an LLM.
- Use read_only and project_ref scoping rather than account-wide access.
- Official · MongoDB
Document databases — the only official NoSQL entry in this roundup.
Document CRUD and aggregation pipelines, plus Atlas management — cluster creation, database users, access lists, performance monitoring — with a --readOnly flag and a disabledTools config for switching off whole categories (atlas, create, update, delete).
Self-hosted npm package; connects to any MongoDB instance or Atlas cluster. · Connection string via env var, Atlas service-account credentials, or OAuth through the Atlas-hosted plugin. · Apache-2.0 · 1,116★
- MongoDB and Atlas only.
- Some features, like the mcpUI web interface, are marked preview.
- Atlas management tools can create billable cloud resources — disable the atlas tool category if you only want data access.
- Community
Postgres-only shops that want performance tuning, not just query access.
SQL execution plus actual DBA tooling: database health checks (indexes, connections, buffer cache, vacuum, replication lag), index tuning with hypothetical-index simulation, and EXPLAIN plan analysis. Two access modes — unrestricted for development, restricted (read-only, time-capped queries) for safer use.
Self-hosted via Docker or pip. · Standard Postgres connection URI passed as the DATABASE_URI environment variable. · MIT · 3,245★
- Postgres only.
- Index tuning and query statistics require the pg_stat_statements and hypopg extensions; without them you get basic SQL and schema browsing only.
- Community project — not backed by the PostgreSQL project or any database vendor.
- Official · Zapier (the integration platform — not a database vendor)
No-code users already on Zapier who want row-level database actions next to their other automations.
Action-based database access for PostgreSQL and MySQL — add row, update row, find rows via lookup columns or custom queries — alongside Zapier's 9,000+ other app integrations from the same endpoint.
Fully managed and hosted by Zapier — nothing to install or run. · Zapier account login; database credentials are stored with Zapier, and you allowlist which actions the AI can call.
- Not a full SQL surface — you get predefined row actions, not an open query console or schema tools.
- Your database credentials are stored with a third party (Zapier).
- Usage is metered under Zapier MCP plan limits.
- Closed-source hosted service — no repo, stars, or license to inspect, unlike everything else on this list.
Databases MCP servers compared
| Server | Maintained by | Runs | Auth | License |
|---|---|---|---|---|
| MCP Toolbox for Databases | Self-hosted: single Go binary, Docker, or npx, with stdio and HTTP transports. A separate managed offering exists, but only for Google Cloud databases. | Connection credentials via environment variables or tools.yaml config; integrated Google Cloud IAM for cloud-hosted sources. | Apache-2.0 | |
| DBHub | Community | Self-hosted via npm or Docker; stdio and HTTP transports; includes a built-in Workbench web UI. | Single DSN connection string with credentials embedded; SSH tunneling and SSL/TLS supported. | MIT |
| Supabase MCP Server | Supabase | Hosted remote endpoint — no local install; project-scoped MCP URLs can be generated from the Supabase dashboard. | OAuth 2.1 browser login on the hosted endpoint (mcp.supabase.com); personal access token for CLI and self-hosted setups. | Apache-2.0 |
| MongoDB MCP Server | MongoDB | Self-hosted npm package; connects to any MongoDB instance or Atlas cluster. | Connection string via env var, Atlas service-account credentials, or OAuth through the Atlas-hosted plugin. | Apache-2.0 |
| Postgres MCP Pro | Community | Self-hosted via Docker or pip. | Standard Postgres connection URI passed as the DATABASE_URI environment variable. | MIT |
| Zapier MCP (PostgreSQL / MySQL) | Zapier (the integration platform — not a database vendor) | Fully managed and hosted by Zapier — nothing to install or run. | Zapier account login; database credentials are stored with Zapier, and you allowlist which actions the AI can call. | — |
How to choose
Pick MCP Toolbox if you run more than one engine, or anything on Google Cloud — it covers the most databases and is the best-maintained option here. Pick DBHub if you want one small server with a DSN, a read-only flag, and nothing else to learn. Pick Postgres MCP Pro if Postgres is your only database and you care about index tuning and query plans, not just query access. Pick your vendor's official server (Supabase, MongoDB, Neon) if your data already lives on their platform — the hosted OAuth endpoints are the least setup. Pick Zapier MCP only if you can live with row-level actions instead of a real SQL surface.
How this ranking works
Every server here is real and was checked against its repository or vendor documentation on the date above — stars and licenses come from the GitHub API, not from memory. Order reflects verifiable criteria only: official support, maintenance recency, adoption, auth model, and documented capability coverage. We do not run paid placements, and we list every caveat we find, including the unflattering ones.
Common questions
- Is there an official MCP server for databases?
- Not for databases as a category — Anthropic's reference Postgres and SQLite servers were archived in 2025 and are no longer maintained. What exists instead: Google's MCP Toolbox (official, multi-engine) and per-vendor officials from Supabase, MongoDB, and Neon (neondatabase/mcp-server-neon, 626 stars, pushed August 31, 2026), all verified active as of September 1, 2026.
- What does setup and auth actually look like?
- For self-hosted databases you run a local server via npx, binary, or Docker and hand it a connection string with credentials embedded — DBHub, Postgres MCP Pro, and MCP Toolbox all work this way, so create a least-privilege database user and turn on the read-only flag where offered. Browser OAuth with zero local install only exists for platform-hosted databases: Supabase at mcp.supabase.com and Neon at mcp.neon.tech.
- Is it safe to point one of these at a production database?
- The vendors themselves say to be careful: Neon's docs recommend development and testing rather than production, and Supabase points you at its security best-practices guide before you connect an LLM at all. Read-only modes — DBHub's flag, MongoDB's --readOnly, Postgres MCP Pro's restricted mode — are the minimum sensible guardrail, and none of them removes the need to review what the model wants to run.
- Why isn't executeautomation/mcp-database-server ranked? It shows up in a lot of lists.
- It's real (380 stars, MIT, covering SQLite, SQL Server, Postgres, and MySQL) but had no push since August 27, 2025 when we checked on September 1, 2026 — more than a year of inactivity. DBHub covers the same engines and is actively maintained, so it took that slot.
MCP servers we build
Automation Squad publishes its own open-source MCP servers and agent skills, documented with the same rules this ranking uses.