Key takeaways
- No single server wins. Vendor servers go deepest on one platform; multi-database servers trade depth for one setup across engines.
- Check the read-only default. DBHub and MCP Toolbox let every test write through until configured. Snowflake's SQL tool and ClickHouse's server start read-only.
- Remote auth varies. Supabase, Neon, PlanetScale and Snowflake host OAuth servers. DBHub and Toolbox ship with auth off but can enable it. Faucet, the open-source database-to-REST-API and MCP server we make, requires an API key on every request.
- Tool definitions cost context: from 207 tokens (Toolbox SQLite) to 1,400 (Faucet), counted with the o200k tokenizer.
- Skip the archived Anthropic Postgres and SQLite servers. See our Postgres MCP server comparison for maintained options.
We compared twelve database MCP servers on the questions that decide a rollout (which databases, who can connect, what stops a write, what it costs) and ran three of them through the same hands-on test.
What is the best MCP server for databases?
The best MCP server for databases depends on where your data lives. For broad open-source coverage, use MCP Toolbox for Databases; for a light local server across SQL engines, DBHub; for Microsoft stacks with entity-level permissions, SQL MCP Server; for Postgres tuning, Postgres MCP Pro. On a single platform, use its own server: Supabase, Neon, PlanetScale, Snowflake or Oracle SQLcl. To self-host one REST API and MCP server across seven SQL engines, Faucet, the open-source database-to-REST-API and MCP server (our product), fits; for enterprise OAuth, DreamFactory; for hosted SaaS plus databases, CData Connect AI.
Database MCP servers compared at a glance
| Server | Best for | Databases | License, pricing, hosting | Remote MCP auth | Read-only control | Audit log | Maturity | REST API too? |
|---|---|---|---|---|---|---|---|---|
| MCP Toolbox for Databases | Broad coverage, Google Cloud | Google Cloud DBs, PostgreSQL, MySQL, SQL Server, Oracle, MongoDB, SQLite and more | Apache-2.0, free; self-hosted (Google also hosts endpoints) | Opt-in OAuth 2.1 (MCP Authorization); off by default | Per tool; none for the SQLite source | Not evaluated | v1.14.0 (Oct 2026) | Not evaluated |
| DBHub | Light local multi-DB server | PostgreSQL, MySQL, MariaDB, SQL Server, Oracle, SQLite | MIT, free; self-hosted | Optional static bearer tokens (--auth-token); off by default | readonly per tool in dbhub.toml | Not evaluated | v1.5.0 | No |
| SQL MCP Server (Data API builder) | Microsoft SQL / Azure | SQL Server, Azure SQL, Cosmos DB, PostgreSQL, MySQL | MIT, free; self-hosted | Entra ID, custom JWT, App Service | Per-entity role actions; no raw SQL | OpenTelemetry traces | MCP since DAB 1.7 | Yes, plus GraphQL |
| Postgres MCP Pro | Postgres tuning | PostgreSQL | MIT, free; self-hosted | Not described | --access-mode=restricted | Not evaluated | Not evaluated | No |
| Supabase MCP | Supabase projects | Supabase Postgres | Apache-2.0; hosted (Supabase account), CLI or self-hosted | OAuth on hosted only | read_only=true | Not evaluated | Not evaluated | Platform has one |
| Neon MCP Server | Neon dev branches | Neon Postgres | MIT; hosted (Neon account) | OAuth or API key | readonly=true, READ ONLY transactions | Not evaluated | Not for production, per Neon | Platform has one |
| PlanetScale MCP | PlanetScale insights | PlanetScale Vitess and Postgres | Hosted in your PlanetScale account | OAuth, service tokens | Org-level none / read-only / read-write (default: read-write) | Not evaluated | Not evaluated | Platform API |
| Snowflake-managed MCP | Snowflake + Cortex | Snowflake | Part of Snowflake; inside your account | Snowflake OAuth or External OAuth | SQL tool read-only by default (read_only: true); RBAC, per-tool grants | Not evaluated | GA since Nov 2025 | Snowflake SQL API |
| Oracle SQLcl MCP | Oracle, local | Oracle Database | Free, per Oracle; local (stdio) | Local only | DB user privileges; restrict levels | DBTOOLS$MCP_LOG table | Since SQLcl 25.2; 26.3 current | No (ORDS is separate) |
| Faucet | One role model, 7 SQL engines | PostgreSQL, MySQL, MariaDB, SQL Server, Oracle, Snowflake, SQLite | MIT, free; self-hosted | API key | GET-only role or service read_only | —* | v0.1.x, pre-1.0 | Yes, with OpenAPI 3.1 |
| DreamFactory | Enterprise OAuth, support | SQL engines, NoSQL (MongoDB, Couchbase), file storage (S3, Azure Blob) | Apache-2.0 core plus commercial subscriptions; self-hosted | OAuth 2.0; opt-in API key | Tools filtered to the role’s services | Not evaluated | Not evaluated | Yes |
| CData Connect AI | Hosted, SaaS + DBs | 350+ sources | Commercial, free to start; hosted | Sign-in required | Granular permissions | Audit trails, per CData’s listing | Not evaluated | Not evaluated |
“Not evaluated” means we did not check it, not that the feature is missing. Vendor facts were checked against each project’s own docs on 2026-10-16. Toolbox, DBHub and Faucet were tested hands-on; the rest are profiled from docs.
* Re-testing on Faucet v0.1.16.
Should you still use Anthropic’s Postgres or SQLite MCP server?
No. Anthropic’s reference servers for Postgres and SQLite now live in modelcontextprotocol/servers-archived, whose README states “NO SECURITY GUARANTEES ARE PROVIDED FOR THESE ARCHIVED SERVERS.” The Postgres one had a read-only bypass via stacked statements that many tutorials still overlook. Pick a maintained server from this list or from our Postgres MCP server comparison.
How we evaluated
We make Faucet, so we tried to keep the comparison checkable. Every non-Faucet fact comes from the project’s own README, docs, changelog or pricing page, re-read on 2026-10-16 and listed under Sources. Faucet facts come from its source code. Drafting was AI-assisted; every vendor fact and test number was checked by hand against the sources and raw results.
We judged each server on five questions: which databases it covers, what a remote client must present, whether read-only exists and is the default, how large its tools/list response is, and where it runs and how it is priced.
Hands-on. On 2026-10-16 we ran MCP Toolbox, DBHub, Faucet and one extra open-source server (ExecuteAutomation’s database-server, not one of the twelve) against the same fixture: the Chinook 1.4.5 SQLite database, a fresh copy per tool and mode. Write verdicts come from reading the file back over a separate read-only connection, not from the server’s reply. Each write path got an INSERT, an UPDATE and a DELETE; the raw-SQL tools of DBHub, Toolbox and ExecuteAutomation also got a CTE-prefixed DELETE (WITH x AS (SELECT 1) DELETE ...). Faucet’s raw-SQL tool was refused in every tested mode, so it got no CTE test; on a writable service with raw SQL allowed it runs any SQL. Faucet REST writes as admin with defaults were not measured. Token counts are the compact tools/list JSON run through OpenAI’s o200k tokenizer (Claude’s tokenizer gives different counts). Cold start is launch to first successful query, median of five runs.
The test covered SQLite only; behaviour on other engines was not measured. Faucet’s query performance across engines is covered separately in our MCP database benchmark.
Hands-on results
| MCP Toolbox 1.14.0 | DBHub 1.5.0 | Faucet (v0.1.15-4-g2327d63) | ExecuteAutomation 1.1.0 | |
|---|---|---|---|---|
| tools/list: tools / bytes / o200k tokens | 2 / 917 / 207 | 2 / 1,481 / 347 (read-only config: 1 / 461 / 112) | 8 / 6,285 / 1,400 | 10 / 1,842 / 404 |
| Annotations | destructiveHint:true on both tools, including list_tables | Full set of hints | —* | None |
| Writes with defaults | execute_sql 4/4 landed | execute_sql 4/4 landed, including the CTE DELETE | —* (raw_sql refused; off by default) | write_query 3/3 landed; read_query refused 4/4 |
| Read-only config | No Toolbox setting. SQLITE_DATABASE=file:...?mode=ro: 0/4 landed (enforced by SQLite) | dbhub.toml with readonly = true (1 file, 7 lines): 0/4 landed | Role with --verbs GET plus API key: 0/9 landed (REST and MCP). Service read_only=true: 0/9 over HTTP even as admin with raw_sql on, 0/6 over stdio | No read-only mode |
| Unauthenticated MCP request | 200 (binds 127.0.0.1) | 200 (auth disabled, binds 0.0.0.0) | 401 | stdio only |
Rows from SELECT * on InvoiceLine (2,240 rows) | 2240 | 2240 | 25 by default, 1000 max (measured via faucet_query and REST, not raw SQL) | 2240 |
| Cold start, median of 5 | 0.114 s (HTTP) | 1.512 s (HTTP) | 0.051 s (HTTP), 0.025 s (stdio) | 0.731 s (stdio) |
* Re-testing on Faucet v0.1.16.
Cold start is not a load or throughput number.
1. MCP Toolbox for Databases — Best for broad open-source coverage, especially on Google Cloud
Best for: teams that need many data sources, or that run on AlloyDB, Cloud SQL, Spanner or BigQuery.
Google’s open-source MCP Toolbox (Apache-2.0, repo now googleapis/mcp-toolbox, formerly genai-toolbox) is a Go binary that serves tools you define in a config file, or prebuilt toolsets per source. Version 1.14.0 shipped on October 8, 2026.
Strengths
- The widest source list among the open-source servers here: Google Cloud databases plus PostgreSQL, MySQL, MariaDB, SQL Server, Oracle, MongoDB, Redis, Neo4j, Snowflake, Trino, SQLite and more.
- Small surface: 2 tools in 207 tokens, and a 0.114 s cold start in our test.
- Both prebuilt tools carry
destructiveHint: true, so annotation-aware clients can ask first. - Spec-style MCP auth: with
mcpEnabled: true, Toolbox acts as an OAuth 2.1 resource server, validating your issuer’s JWTs and publishing/.well-known/oauth-protected-resource. Google also runs managed MCP endpoints for AlloyDB, Cloud SQL, Spanner, BigQuery, Firestore and Bigtable.
Things to know
- With the SQLite prebuilt, all 4 of our writes landed. The SQLite source has no read-only option; a
mode=roconnection URI made SQLite refuse them. - MCP auth is opt-in: with defaults,
/mcpanswered an unauthenticatedtools/listwith 200. It binds to 127.0.0.1 by default, which limits exposure. list_tablesis annotated destructive too, so clients may prompt for harmless reads.- No default row cap:
SELECT *returned all 2,240 rows.
Who should choose it
Platform teams standardising on one open-source server across many engines who are willing to write tool configs, and Google Cloud teams that may want managed endpoints or OAuth later.
2. DBHub — Best for a lightweight local multi-database server
Best for: developers who want an agent in their IDE talking to several SQL databases with almost no setup.
DBHub from Bytebase (MIT, TypeScript) covers PostgreSQL, MySQL, MariaDB, SQL Server, Oracle and SQLite and exposes two tools by default, execute_sql and search_objects, with opt-in extras.
Strengths
- Tiny context cost: 347 tokens by default, 112 with read-only configured.
- Read-only works at the statement level: with
readonly = trueindbhub.toml, 0 of 4 writes landed, including the CTE-prefixed DELETE. - Per-tool
max_rowslimits and custom parameterised tools in the same TOML file. - Runs over stdio or HTTP. Over HTTP,
--auth-tokenrequires bearer tokens on every request (401 without one), and--allowed-hostsadds DNS-rebinding protection.
Things to know
readonlydefaults to false andmax_rowsto unlimited. All 4 of our writes landed with defaults.- Tool options such as
readonlyandmax_rowscan only be set in TOML, not with command-line flags. - HTTP auth is off by default and
--hostdefaults to 0.0.0.0. In our run the server logged “Auth: disabled” and returned 200 to an unauthenticatedtools/list. Set--auth-tokenor bind to localhost. - Slowest cold start in our set: 1.512 s median over HTTP.
Who should choose it
Developers who want one small server for several SQL engines in their IDE, and small teams willing to manage static tokens for shared HTTP use.
3. SQL MCP Server (Data API builder) — Best for Microsoft SQL and Azure teams that want entity-level permissions
Best for: SQL Server and Azure SQL shops that want agents to do CRUD through a governed contract rather than free-form SQL.
SQL MCP Server is part of Microsoft’s Data API builder (DAB) from version 1.7. DAB is MIT-licensed and, in Microsoft’s words, “truly free, with no premium tier”. The same JSON config serves REST, GraphQL and MCP.
Strengths
- Seven DML tools (describe, create, read, update, delete, execute and aggregate), plus stored procedures as custom tools in 2.0.
- No NL2SQL and no DDL by design: agents can only touch entities you configured.
- Per-entity, per-role actions with field include and exclude, applied to REST, GraphQL and MCP alike.
- Entra ID, custom JWT and App Service auth, OpenTelemetry tracing, health checks and caching.
Things to know
- DAB does not list SQLite, Oracle or Snowflake among its supported databases.
- Each table, view or procedure must be declared as an entity, which is more setup than pointing a server at a database.
- The default auth provider is “Unauthenticated”, meant for running behind a trusted front end; set a real provider before exposing it.
- No ad-hoc SQL means exploratory questions need entities or aggregates defined first.
Who should choose it
Azure or SQL Server teams that value a deterministic, permissioned tool surface over ad-hoc queries. Our SQL Server MCP server comparison goes deeper.
4. Postgres MCP Pro — Best for Postgres performance tuning and health checks
Best for: engineers using an agent as a Postgres DBA assistant.
Postgres MCP Pro from Crystal DBA (MIT, Python) targets PostgreSQL only, tested on versions 15 to 17.
Strengths
- Nine tools including
explain_querywith hypothetical indexes,get_top_queriesfrompg_stat_statements, workload index analysis andanalyze_db_health. --access-mode=restrictedlimits SQL to read-only transactions with execution-time limits.
Things to know
- Postgres only.
- Unrestricted mode allows full read/write and schema changes; pick restricted for anything shared.
- Transports are stdio and SSE; the README does not describe client authentication for SSE.
Who should choose it
Postgres teams diagnosing slow queries and missing indexes who will run it in restricted mode. Compare it with other options in our Postgres MCP server roundup.
5. Supabase MCP — Best for building on Supabase projects
Best for: developers whose database is a Supabase project.
Supabase’s MCP server (Apache-2.0) runs hosted at https://mcp.supabase.com/mcp and covers far more than SQL: projects, migrations, logs, edge functions and docs.
Strengths
- Hosted with a browser login, so there are no database credentials in a config file.
read_only=truedrops mutating tools;project_refscopes to one project;featureslimits tool groups.
Things to know
- Supabase Postgres only.
- The local CLI and self-hosted versions offer “a limited subset of tools and no OAuth 2.1”.
- Supabase’s README points to its security guidance on connecting an LLM to your projects; read it before use.
Who should choose it
Supabase teams that want an agent to manage the whole project, not only run SQL; scope it with project_ref and read_only=true. Weighing Supabase against a self-hosted API layer? See Faucet vs Supabase.
6. Neon MCP Server — Best for branch-based schema changes on Neon during development
Best for: developers on Neon who want agents to test migrations on a throwaway branch.
Neon’s server (MIT) is hosted at https://mcp.neon.tech/mcp over Streamable HTTP.
Strengths
prepare_database_migrationapplies changes to a temporary branch;complete_database_migrationapplies them to main.- OAuth by default or a Neon API key;
?readonly=truehides write tools and runsrun_sqlin aREAD ONLYtransaction. roleNameruns SQL as a specific Postgres role, andprojectIdscopes access.
Things to know
- Neon’s README says: “We do not recommend using the Neon MCP Server in production environments.”
- Neon Postgres only.
Who should choose it
Neon users who want agents to rehearse migrations on a branch during development; not for production, per Neon.
7. PlanetScale MCP — Best for query insights and replica reads on PlanetScale
Best for: PlanetScale customers who want agents to investigate performance.
PlanetScale’s hosted server at https://mcp.pscale.dev/mcp/planetscale covers Vitess and Postgres databases with 26 tools.
Strengths
- Reads route to a replica by default when the branch has replicas.
- Organisation admins choose none, read-only, or read and write for MCP queries.
- The write tool blocks
UPDATEorDELETEwithoutWHERE, blocksTRUNCATE, and requires human confirmation for DDL. - An insights-only endpoint omits query execution entirely.
Things to know
- Organisation MCP query access defaults to “Read and write queries”; lower it if agents should only read.
- PlanetScale databases only; the docs page does not list MCP-specific pricing.
- PlanetScale itself advises caution before giving LLMs write access to production.
Who should choose it
PlanetScale teams who want an agent to investigate slow queries on replicas without write risk; use the insights-only endpoint if you don’t need query execution.
8. Snowflake-managed MCP server — Best for Snowflake accounts using Cortex
Best for: data teams that want agents governed by the same Snowflake roles as everyone else.
Snowflake’s managed MCP server has been generally available since November 4, 2025, and runs inside your account.
Strengths
- Tools for Cortex Agents, Cortex Search, Cortex Analyst (semantic views),
SYSTEM_EXECUTE_SQLand custom UDFs or procedures. SYSTEM_EXECUTE_SQLis read-only by default (read_onlydefaults totrue).- Snowflake OAuth or External OAuth, with grants needed per tool, not just on the server.
Things to know
- Tools only: no resources, prompts or sampling. Up to 50 tools per server, and SQL tool responses are truncated at 250 KB.
- Snowflake advises against exposing
SYSTEM_EXECUTE_SQLon the same server as agents built on semantic views, since it lets clients bypass them. - No dynamic client registration, which some MCP clients expect. Snowflake only.
Who should choose it
Snowflake teams whose agents should work through Cortex and semantic views under existing roles. Teams that also need Postgres or SQL Server behind the same agent will need a second server.
9. Oracle SQLcl MCP Server — Best for Oracle developers working locally
Best for: Oracle developers who already use SQLcl or SQL Developer for VS Code.
Oracle describes SQLcl’s MCP server as free. It runs locally over stdio and connects through saved SQLcl connections. The MCP server arrived in SQLcl 25.2 (July 2025); 26.3.0 is the current release.
Strengths
- A small tool set that has grown with each release:
list-connections,connect,disconnect,run-sql,run-sqlclandschema-information, plus async queries (25.4) andskills_sync,annotation_generateandcontext_retrieve(26.3). - Every request is logged to a
DBTOOLS$MCP_LOGtable, including client and model, andV$SESSIONshows the MCP client and model inMODULEandACTION.
Things to know
- The restrict-level default changed. Oracle’s 26.3 docs still describe level 4 (no host commands or scripts) as the default, but the SQLcl 26.2 changelog says “SQLcl MCP startup now defaults to unrestricted.” Pass
-R 4explicitly. - Restrict levels limit SQLcl features, not DML: what the agent can change is set by the database user’s privileges, so connect as a read-only user.
- Local stdio only; Oracle offers OCI Database Tools MCP and an ORDS MCP endpoint for remote use.
Who should choose it
Oracle teams starting with local agent use who want a database-side audit trail of every agent request.
10. Faucet — Best for self-hosting one REST API and MCP server across seven SQL engines
Best for: teams that want agents and applications on the same API, with one role model, across PostgreSQL, MySQL, MariaDB, SQL Server, Oracle, Snowflake and SQLite.
Faucet (MIT, Go) is a single binary that introspects a database and serves a REST API, an OpenAPI 3.1 document, an admin UI and an MCP server at /mcp on one port. A read-only agent setup:
faucet db add --name chinook --driver sqlite --path "$PWD/chinook.db"
faucet role create --name readonly --verbs GET
faucet key create --role readonly
faucet serve
claude mcp add --transport http faucet http://localhost:8080/mcp \
--header 'X-API-Key: YOUR_API_KEY'
Use an absolute SQLite path; a relative one resolves against the directory faucet serve starts in.
Strengths
- One role model for REST and MCP: with a GET-only key, 0 of 9 writes landed.
- The service
read_onlyflag let 0 of 9 writes through over HTTP, even for an admin with raw SQL enabled (0 of 6 over stdio). - Raw SQL is off by default. It runs only when the service allows raw SQL (set in the admin UI or REST API), the service is not read-only, and, for API keys, the role grants all five verbs on
_sql. - Unauthenticated data API and
/mcprequests got 401.faucet_queryand REST return 25 rows by default, up to 1,000. - The same tables are a REST API with an OpenAPI 3.1 document, which suits apps and Custom GPT Actions; see our database-to-REST-API tools comparison and the Faucet MCP server and RBAC docs.
Things to know
- A newer project. Faucet is at v0.1.x, so it has a shorter track record than the vendor servers above.
- API-key auth. Clients send an
X-API-Keyheader; there is no OAuth sign-in flow, so OAuth-only clients connect through a Custom GPT Action or the REST API instead. - SQL databases only, REST rather than GraphQL. Roles are scoped by service, table and verb.
Who should choose it
Self-hosting teams on several SQL engines that want one API for apps and agents, with API-key auth. Engine guides: PostgreSQL REST API and MCP server, SQL Server REST API and MCP server.
11. DreamFactory — Best for enterprises that need OAuth for AI clients and commercial support
Best for: larger organisations exposing many databases and file stores to AI clients under one governed API platform.
DreamFactory has a built-in MCP server over Streamable HTTP that exposes the database and file storage services you select, with tools prefixed per API.
Strengths
- AI clients such as VS Code, Claude and Cursor authenticate with OAuth 2.0; headless clients can use an API key if the service opts in.
- Tools are filtered to the services the connecting user’s role allows.
- Connectors span many SQL engines, NoSQL stores such as MongoDB and Couchbase, and file storage like S3 and Azure Blob.
Things to know
- DreamFactory’s pricing page lists its free Open Source plan as PostgreSQL only, and its GitHub README names SQL Server, Oracle and Snowflake as commercial connectors.
- Its docs and pricing page do not state which editions include the MCP server; check with DreamFactory.
- Heavier to run than a single binary.
Who should choose it
Enterprises that need OAuth sign-in for AI clients, connectors beyond SQL, and a vendor support contract.
12. CData Connect AI — Best for hosted access to databases and SaaS from one remote server
Best for: teams whose agents need SaaS data (CRM, ERP, finance) alongside SQL databases, without running servers.
CData Connect AI is a commercial, fully hosted remote MCP server listed in Claude’s connector directory.
Strengths
- 350+ sources, read and write, from one endpoint.
- Granular permissions and audit trails, according to its connector listing.
Things to know
- Commercial and hosted only, so queries run through CData’s service. Plans are subscriptions, and CData says you can start free.
Who should choose it
Teams whose agent questions span SaaS apps and databases, and who accept routing data through a hosted vendor to avoid running servers.
Which database MCP server should you use for your database?
PostgreSQL
Postgres MCP Pro in restricted mode for tuning; Supabase’s or Neon’s server on those platforms. More options in our Postgres MCP server comparison.
SQL Server and Azure SQL
SQL MCP Server for entity-level CRUD with no raw SQL; Toolbox, DBHub and Faucet also connect. See our SQL Server MCP server comparison.
Oracle
SQLcl’s MCP server locally, with -R 4 and a read-only database user; OCI Database Tools MCP or ORDS for remote access.
Snowflake
The managed server keeps agents inside account RBAC and OAuth. Faucet and Toolbox also reach Snowflake alongside other engines.
Several engines at once
Toolbox for coverage, DBHub for footprint, Faucet for a REST API under the same roles. Our open-source multi-database MCP server comparison tests them head to head.
Also considered
- MongoDB MCP Server (Apache-2.0): official, for MongoDB and Atlas, with
--readOnly(off by default). Left out because this list covers SQL engines. - AWS Labs MCP servers (Apache-2.0): stdio servers for Aurora, DSQL, DynamoDB, Redshift and more. Their postgres-mcp-server had CVE-2026-87911, a read-only bypass via
COPY ... TO PROGRAM, fixed in 1.1.7; covered in our Postgres comparison. - ClickHouse mcp-clickhouse (Apache-2.0): read-only by default; writes need
CLICKHOUSE_ALLOW_WRITE_ACCESS=true. Left out to keep the list at twelve. - MindsDB: a federated query engine. Its MCP docs URL redirected on 2026-10-16, so we could not verify current details.
Vendor-native vs vendor-neutral database MCP servers
Vendor-native servers (Supabase, Neon, PlanetScale, Snowflake, Oracle) know their platform, and most hosted ones ship OAuth. The cost is one server and permission model per platform, a pattern we covered in every database vendor shipping its own MCP server. Vendor-neutral servers (Toolbox, DBHub, DAB, Faucet, DreamFactory, CData) put one policy layer in front of several engines; of the self-hosted ones, Toolbox and DreamFactory support OAuth, DBHub uses static bearer tokens and Faucet API keys.
For a head-to-head of DBHub and MCP Toolbox on read-only, auth and cold start, see our open-source multi-database MCP server comparison.
Bottom line
Use the vendor’s own server when you live on one platform; Toolbox or DBHub for a quick multi-database setup, with read-only and auth configured; SQL MCP Server for a Microsoft stack; DreamFactory or CData for enterprise breadth; and Faucet for apps and agents on one self-hosted API, if API-key auth is acceptable. Whatever you choose, give your agent read-only database access first.
To try Faucet, follow the getting started guide, then connect Claude to your database for client setup.
Frequently asked questions
What is a database MCP server?
A database MCP server is a process that exposes a database to AI clients through the Model Context Protocol. It publishes tools such as list tables, describe table and run query, and clients like Claude Code, Cursor or VS Code call those tools instead of connecting to the database directly. The server decides what the agent can see and change, so its auth and read-only controls matter as much as its features.
Which MCP server supports the most databases?
Among the servers in this comparison, CData Connect AI lists 350+ sources, including SaaS applications, as a hosted commercial service. Google's MCP Toolbox for Databases is the broadest open-source option, covering Google Cloud databases plus PostgreSQL, MySQL, SQL Server, Oracle, MongoDB, Redis, Snowflake and more. DBHub and Faucet each cover six or seven SQL engines.
Are database MCP servers safe to use in production?
Some vendors say not yet: Neon's README states it does not recommend its MCP server in production. For any server, use a database user with minimal privileges, turn on the server's read-only mode, require authentication on HTTP endpoints, and cap result sizes. Our database MCP server security checklist covers each step.
How many tokens do MCP tool definitions use?
In our harness, the full tools/list response cost 207 tokens for MCP Toolbox's SQLite prebuilt, 347 for DBHub, 404 for ExecuteAutomation's server and 1,400 for Faucet (OpenAI's o200k tokenizer; Claude's tokenizer gives different counts). None of them grew with table count. Servers with dozens of tools cost more; see the MCP context window tax.
Can one MCP server connect to multiple databases?
Yes. MCP Toolbox, DBHub, SQL MCP Server (Data API builder), Faucet, DreamFactory and CData Connect AI can all expose several databases through one server. They differ in how access is scoped: per tool, per entity, per role or per user. Our open-source multi-database MCP server comparison goes deeper on the self-hosted ones.
Which database MCP servers work with Claude Code, Cursor and ChatGPT?
Servers using stdio, SSE or Streamable HTTP work with Claude Code, Cursor and VS Code. ChatGPT connectors sign in with OAuth, so hosted OAuth servers such as Supabase, Neon and PlanetScale fit there. Faucet authenticates with API keys; use it from ChatGPT through a Custom GPT Action that imports its OpenAPI document, or from the OpenAI Responses API's MCP tool with an API-key header. Both need HTTPS, so put Faucet behind a TLS reverse proxy. See connect ChatGPT to your database.
Sources
- googleapis/mcp-toolbox README (checked )
- MCP Toolbox documentation (checked )
- MCP Toolbox docs: Toolbox with MCP Authorization (checked )
- Google Cloud: supported products for remote MCP servers (checked )
- bytebase/dbhub README (checked )
- DBHub docs: TOML configuration (checked )
- DBHub docs: command-line options (checked )
- Microsoft Learn: SQL MCP Server overview (checked )
- Microsoft Learn: What is Data API builder? (checked )
- crystaldba/postgres-mcp README (checked )
- Supabase MCP README (checked )
- neondatabase/mcp-server-neon README (checked )
- PlanetScale MCP server docs (checked )
- Snowflake-managed MCP server docs (checked )
- Snowflake release note: managed MCP server GA (Nov 4, 2025) (checked )
- Oracle MCP Servers (checked )
- Oracle SQLcl changelog (checked )
- Oracle docs (SQLcl 26.3): configuring restrict levels for the SQLcl MCP Server (checked )
- Oracle docs (SQLcl 26.3): monitoring the SQLcl MCP Server (checked )
- DreamFactory MCP server docs (checked )
- DreamFactory GitHub repository (checked )
- DreamFactory pricing (checked )
- CData Connect AI in the Claude connector directory (checked )
- CData Connect AI product page (checked )
- executeautomation/mcp-database-server README (checked )
- mongodb-js/mongodb-mcp-server README (checked )
- awslabs/mcp README (checked )
- AWS Security Bulletin 2026-104 (CVE-2026-87911) (checked )
- ClickHouse/mcp-clickhouse README (checked )
- modelcontextprotocol/servers-archived (checked )
- Faucet source code (v0.1.15-4-g2327d63) (checked )
Last reviewed