Bonnard
On this page

This page compares 23 database MCP servers, our own included. An MCP server is a service that lets Claude or ChatGPT query your database. Each entry says who it’s for, how people sign in, whether it can change your data, which AI apps it works in and what it costs, and links each fact to its source. The table lists all of them, and the next section says which we’d pick for your situation.

We sell one of these, so our entry sits apart from the numbered list and says where we fall short.

Database MCP servers compared, checked 16 Nov 2026
OptionWho hosts itWho signs inCan it write to your database?Charts in chatPrice
Bonnard (ours)We doEach person, own accountNo, except SQL Server, where your login decidesYes, in AI apps with MCP AppsFree for 3,000 calls a month (about 600 questions), then $0.95 per 100 if you turn on extra usage
BigQuery MCP server (Google)GoogleEach person, own accountYes, via its main SQL tool; a second tool is read-onlyNot documentedNo MCP charge; queries billed as normal
Snowflake MCP server (Snowflake-managed)SnowflakeEach person, own accountSQL tool only if an admin allows; procedure and agent tools canNot documentedNo MCP charge; AI billed per token, plus warehouse time
Databricks Genie One MCPDatabricksEach person, own accountNoYes, in AI apps with MCP AppsGenie AI free for people (not apps) until 31 Jan 2027; warehouse billed
ClickHouse MCP servers (Cloud and self-hosted)ClickHouse, or youCloud: each person. Self-hosted: one shared loginCloud: no. Self-hosted: only if you allow itNot documentedNo MCP charge; queries billed as normal
MotherDuck MCP serverMotherDuck, or youEach person, own accountYes (query_rw); read-scaling tokens (Business, Enterprise) block writesYes (Dives)On all plans; Lite $0 a month
PostHog MCP serverPostHogEach person, own accountYes, until you turn on read-only modeIn some AI appsNo MCP charge; queries count toward your PostHog bill
Supabase MCP serverSupabase, or youEach developer, own accountYes, until you turn on read-only modeNot documentedNo MCP price listed
Postgres MCP server: Postgres MCP ProYouOne shared login for everyoneYes, until you start it in restricted modeNoFree, open source (MIT licence)
MySQL MCP server: mcp-server-mysqlYouOne shared login for everyoneNo, unless you turn on each kind of writeNoFree, open source (MIT licence)
SQL Server MCP: Microsoft’s SQL MCP ServerYouEach person with Entra ID, or no one (the default)Yes; all tools on, limited by roles in its configNoFree
Redshift MCP server (AWS Labs)You, on your own machineEach person, with their AWS loginNo, unless you turn on read-write modeNoFree
Google MCP Toolbox for DatabasesYouOne shared database login; optional per-person token checkDepends on your tools; read-only mode covers 4 Google Cloud databasesNo built-in chartsFree
DBHubYouOne shared login for everyoneYes, until you set readonlyNoFree, open source (MIT licence)
Metabase MCP serverMetabase, or youEach person, own accountYes, unless you connect a read-only database userYes, once an admin enables each AI appOn all plans; Starter $100 a month for 5 users
Looker MCP serverGoogleEach person, own accountAn admin turns each tool onNot documentedNo MCP charge; uses Looker’s API limit (1,000 a month on Standard). Looker by quote
Power BI: Fabric IQ MCPMicrosoftEach person, own accountNoNot documentedNo MCP price; Power BI Pro $14 per user a month
Hex MCP serverHexEach person, own accountEdits Hex project draftsIn Claude, for 24 hoursTeam plan, $75 per editor a month, plus Hex credits
Cube MCP serverCubeEach person, own accountEdits Cube’s model files, by roleIn some AI appsFree tier, then from $40 per developer a month
dbt MCP serverdbt, or youEach person, or an access tokenYes. execute_sql doesn’t block writesNot documentedMCP on all plans; metrics tools need Starter, $100 per user a month
BasedashBasedashEach person, own accountRead-only by default; writes only in Basedash’s web app, with approvalNo, built in BasedashFrom $1,000 a month for up to 25 users, plus AI usage
QuerioQuerioEach person, own accountQuerio advises a read-only loginNo, kept in Querio notebooksFree tier; Core $2,499 a month
BlazeSQLBlazeSQLEach person, own accountNo. Its SQL tool is read-onlyYesTeam Advanced, $800 a month for 3 users

Which database MCP server should you use?

The best MCP servers for databases suit different situations. Pick by where your data lives and who needs to ask.

  • Your data is in BigQuery, Snowflake, Databricks, ClickHouse Cloud, MotherDuck or PostHog, and your questions don’t need data from anywhere else: we’d use that vendor’s own server.
  • Your metrics are already modelled in Looker, Power BI, Cube or Metabase: use that tool’s server, so the numbers match your dashboards. The exception is if you plan to replace the BI tool.
  • You’re a developer working on one database in Claude Code or Cursor: an open-source server you install on your own computer is enough. Turn on its read-only mode and give it a read-only database user.
  • Your data is in Postgres, MySQL, SQL Server or Redshift, and people across the company want to ask it questions from Claude or ChatGPT: this is what we built Bonnard for. One server sits in front of all your databases, each person signs in as themselves, and the agent (Claude or ChatGPT) answers from the descriptions and metrics your team publishes (how it works, context).
  • Your own customers will ask questions about their data through your product: first check how the server keeps each customer’s rows away from the others (tenancy). We support row-level security, database rules that limit which rows each user can see, on Postgres, Redshift and SQL Server (tenancy docs). Cube and Metabase document their own ways of keeping customers apart.

How we compared database MCP servers

We read each option’s docs, pricing page and repository in November 2026. We haven’t run the same tasks against all 23, so this compares what each vendor documents. Vector databases, document stores and integration platforms that connect an assistant to many apps are left out.

Which AI apps can use a database MCP server?

Neither Claude nor ChatGPT connects to a database by itself. An MCP server gives them tools, such as running a query or writing data. MCP is an open protocol, so a server isn’t tied to one assistant (MCP). Where a server works depends on how you reach it. A hosted server that signs each person in with their own account (OAuth) works in any AI app that supports hosted MCP servers, within that app’s plan rules. A server you run on your own machine works only in AI apps installed on that machine, such as Claude Desktop, the ChatGPT desktop app and coding tools (Claude help, OpenAI docs).

Charts are a separate question. A server can send back an interactive chart or dashboard through MCP Apps, an official MCP extension for interactive views that the AI app draws inside the conversation (MCP Apps). The AI app has to support it as well, and one that doesn’t shows the server’s text answer instead.

AI apps and MCP, checked 13 Nov 2026
AI appHosted servers with OAuthServers on your machineCharts (MCP Apps)Source
Claude (web, Desktop, mobile)Yes, on every plan. Free allows one custom connector (Claude’s name for an MCP server you add by its URL), and on Team and Enterprise owners add themDesktop onlyYesClaude help, interactive connectors
Claude CodeYesYesNodocs
ChatGPTOn the website. Full support on Business, Enterprise and Edu, where admins add servers (on Enterprise and Edu they can authorise members too). Pro gets read and fetch tools onlyDesktop app onlyYesOpenAI help, plugin docs
CursorYesYesYes, since version 2.6docs
VS Code with GitHub CopilotYes. On Copilot Business and Enterprise an organisation policy decides whether MCP is onYesYes, experimentaldocs
Microsoft 365 CopilotYes, through an agent your admin lets you uploadNot documentedYesMicrosoft docs
Gemini appYes, for personal accounts in the US (18 or over, added on the web)Not documentedNot documentedGemini help
OpenAI CodexYesYesNot documenteddocs
GooseYesYesYes in the desktop app, experimentaldocs

On Claude Team and Enterprise, owners can set each tool to always allowed, needs approval or blocked, and that setting is how you keep a server read-only in Claude (Claude help).

In ChatGPT you add a hosted server yourself from the Plugins page on the website, or install a ChatGPT plugin a vendor has published, which adds its server for you (OpenAI help). Snowflake’s guide still walks you through developer mode, a ChatGPT setting OpenAI no longer requires.

Where Bonnard fits

Bonnard (ours, not ranked)

We built Bonnard for teams who want to ask questions from Claude or ChatGPT without handing an agent a login that can write. It fits best when your data is in Postgres, MySQL, SQL Server or Redshift.

  • Sign-in: Each person signs in with their own account, through OAuth. An access token lasts 15 minutes and only works with our server’s URL, so a leaked one is soon useless. The AI app renews it for up to 30 days, then the person signs in again (OAuth docs).
  • Can it change data: No, except on SQL Server. Everywhere else we refuse anything but a single SELECT. On SQL Server we add no check of our own and the login’s permissions decide, so connect us with a read-only login there (SQL Server docs).
  • Works in: Any AI app that can add a hosted MCP server with OAuth sign-in. Our setup guide covers Claude, Claude Code, ChatGPT, Cursor, Codex, Microsoft 365 Copilot and Gemini (client setup). Charts and dashboards appear in the chat in AI apps that support MCP Apps, such as Claude and ChatGPT. Other AI apps get a text summary and the first rows (compatibility).
  • Price: The Free plan includes 3,000 calls a month, about 600 questions. A call is one tool call the agent makes to our server, and a question takes about five. Past 3,000, calls stop unless you turn on extra usage at $0.95 per 100 calls, billed in blocks of 1,000 (billing docs). Team is $150 a month with 20,000 calls, then $0.75 per 100. No plan charges per user (pricing).

We host one MCP server in front of the databases you already run and connect from one fixed IP address, 34.187.7.127, so you only add that one address to your database firewall (IP addresses). Ask the agent to delete a row and, on every database except SQL Server, it gets back “Only read-only SELECT queries are allowed.” (errors). Your team writes the descriptions, joins and metrics the agent reads and publishes them as versions. The agent keeps reading the live version until someone publishes a new one (context), and you can roll back to an earlier version in our web console (publishing). The agent sees the first 250 rows of a result, and charts and standard exports can use up to 5,000. A bulk export goes up to 500,000. In Claude Desktop and Claude Code we keep each reply under 48 KB, so a large result gives the agent fewer rows to read (limits).

We don’t support Snowflake or Databricks, and if that’s where your data is, their own servers are the better choice anyway. We also connect to BigQuery, ClickHouse, MotherDuck and PostHog, though for those we’d usually point you to the vendor’s own server first. Each database must be reachable from the internet (connect data), and deploying us inside your private network (VPC deployment) is Enterprise only (pricing). The agent reads your metric definitions, but nothing forces its SQL to follow them, so it can still calculate a metric its own way (metrics). Automated tests of the agent’s answers (evals) aren’t live yet (our context page). We’re not a listed ChatGPT plugin yet, so you add us as a custom MCP server on ChatGPT’s website, which needs Business, Enterprise or Edu for full support (OpenAI help). Setup for each database is in our docs.

Hosted by your database’s vendor

If your data already lives with one of these vendors, start here. They host the server and sign each person in with their own account, so the database’s own permissions follow every question. Read-only behaviour is where they differ most.

1. BigQuery MCP server (Google)

Google’s hosted server lets anyone an admin has given the MCP Tool User role and BigQuery access query it from Claude, ChatGPT or Gemini CLI, signed in as themselves.

  • Sign-in: Each person, with their Google identity. Google Cloud IAM decides what they can see (docs).
  • Can it change data: Yes. Its execute_sql tool can change rows (INSERT, UPDATE, DELETE) and tables (CREATE, ALTER, DROP). execute_sql_readonly is the SELECT-only tool, and an IAM deny policy can block execute_sql (tool reference, docs).
  • Works in: AI apps that support hosted MCP servers. For Claude, ChatGPT and Cursor you create your own OAuth client ID and secret in Google Cloud. Google documents Claude, ChatGPT on the Business plan, Gemini CLI and others (setup guide).
  • Price: Google lists no separate charge for the server. Queries bill as normal BigQuery jobs, and the first 1 TiB processed each month is free (BigQuery docs).

If your data is already in BigQuery, we’d try this before anything else here. It has been out of preview since April 2026 (release notes), and because people sign in as themselves, the permissions you’ve already set up in IAM carry over without extra work.

Check the tool list before anyone connects. The BigQuery listing in Claude’s connector directory shows execute_sql and not the read-only tool, so a team adding it from there gets the tool that writes unless the deny policy is in place first. A query returns at most 3,000 rows and, by default, is stopped after three minutes (docs).

2. Snowflake MCP server (Snowflake-managed)

Snowflake’s managed MCP server is the route Snowflake documents for Claude, ChatGPT and Cursor. An admin creates the server in Snowflake and picks which tools it offers, and each person then signs in as their own Snowflake user.

  • Sign-in: Each person, as their own Snowflake user. Setting up OAuth needs the ACCOUNTADMIN role by default, so your account admin has to do it (docs, security integration).
  • Can it change data: It can, if an admin allows it. The SQL tool is read-only by default, but an admin can change that. Tools that run stored procedures or Cortex Agents (Snowflake’s AI agents, which plan the work and call tools themselves) are marked as able to make changes (docs).
  • Works in: AI apps that support hosted MCP servers and accept the OAuth client ID and secret you create in Snowflake. Snowflake documents Claude and ChatGPT (docs).
  • Price: No separate MCP charge. Cortex Agents bill AI credits per million tokens, at $2.00 per AI credit with global routing or $2.20 regional, on top of warehouse compute for the SQL they run (Cortex pricing).

Most of its value comes from semantic views, which Snowflake describes as objects that store your business metrics and how your entities relate, directly in the database (semantic views). If you’ve built them, the agent answers from your definitions instead of guessing at joins. If you haven’t, expect to build them first.

Snowflake’s recommended setup is to give Claude a Cortex Agent as its tool, rather than Cortex Analyst (Snowflake’s service that turns a question into SQL using your semantic views) or raw SQL, and in August it told customers to move from Cortex Analyst to Cortex Agents (release note). Assistants like Claude connect through MCP, and your own software uses Snowflake’s agent:run API (Cortex Agents). Snowflake warns that if you include the SQL tool, Claude or ChatGPT can query tables directly and skip your semantic views, so their numbers may not match your definitions (docs).

3. Databricks Genie One MCP

Genie One MCP puts Genie, the Databricks interface built for business users, into Claude, ChatGPT and Cursor. Databricks recommends it over its other MCP servers for questions in plain language.

  • Sign-in: An admin sets up Databricks’ Genie plugin in Claude, ChatGPT or Cursor, then each person signs in with their Databricks account. The permissions you’ve set in Unity Catalog, Databricks’ permissions layer, apply to every question (docs).
  • Can it change data: No, Genie One’s tools are read-only (trust docs). Databricks’ separate SQL MCP server, still in beta, permits writes by default (SQL MCP).
  • Works in: Any AI app that supports hosted MCP servers. Databricks publishes its own Genie plugins for Claude, ChatGPT and Cursor (docs).
  • Price: Until 31 January 2027, the AI part of Genie is free for people signed in as themselves. SQL warehouse compute is billed as usual (cost docs).

When a person asks Genie a question before February 2027, Databricks covers the language-model cost and you pay only for the warehouse that runs the SQL. From 1 February 2027 each person gets 150 DBUs (Databricks Units, the unit Databricks bills by) of Genie a month free, worth about $10.50 at Databricks’ list price in US East, and pays $0.07 per DBU after that (Genie pricing).

The free period doesn’t cover service principals, the non-human identities Databricks uses when an app or a scheduled job calls Genie as itself. Those are billed from the first question (cost docs), so if you plan to run Genie from a pipeline or put it behind your own app, budget for that now. If your own software needs to call Genie directly, without MCP, Databricks points to the Genie One API (docs).

4. ClickHouse MCP servers (Cloud and self-hosted)

ClickHouse has two: a hosted server for ClickHouse Cloud, and mcp-clickhouse, an open-source server you run yourself against any ClickHouse.

  • Sign-in: Cloud: anyone in your organisation who can use the SQL Console (ClickHouse Cloud’s in-browser query editor) for that database, signed in with their own ClickHouse Cloud login, once someone with the manage-MCP permission turns it on for that service (docs, setup guide). Self-hosted: everyone shares one ClickHouse login (README).
  • Can it change data: Cloud: no, every tool is read-only (docs). Self-hosted: read-only by default, enforced by ClickHouse itself, until you start it with the setting CLICKHOUSE_ALLOW_WRITE_ACCESS turned on (README).
  • Works in: Cloud: any AI app that supports hosted MCP servers. ClickHouse documents Claude, and points ChatGPT web users to the hosted server (hosted guide). Self-hosted: AI apps on your own computer, including Claude Desktop, Cursor and the ChatGPT desktop app (setup page).
  • Price: Cloud: no separate MCP charge; queries use the service’s compute (pricing). Self-hosted: free and open source under the Apache 2.0 licence (GitHub).

The hosted server isn’t only for admins. Once someone turns it on, anyone with SQL Console access can connect, and since nothing it exposes can write, it’s the one we’d hand a wider team. The self-hosted server suits ClickHouse you run yourself, but its README says the role it sets for each request won’t keep one customer’s data away from another, and that its check for destructive statements can’t be relied on to stop them. Use a read-only ClickHouse user with it.

If you self-host, note that the query tool was renamed from run_select_query to run_query in version 0.3.0, and run_query can write once you allow it (pull request #93). ClickHouse’s comparison table still shows the old SELECT-only name.

5. MotherDuck MCP server

MotherDuck’s hosted server is meant for everyone in your organisation who uses MotherDuck, and it can answer with a Dive: an interactive chart that runs inside Claude or ChatGPT.

  • Sign-in: Each person, with OAuth in most AI apps, or an access token you create in MotherDuck (docs).
  • Can it change data: Yes. Alongside the read-only query tool there’s query_rw, which writes. A read-scaling token, a kind of access token on the Business and Enterprise plans, connects to read-only replicas, so writes are refused (read-only guide, pricing).
  • Works in: Any AI app that supports hosted MCP servers. MotherDuck documents Claude, ChatGPT, Cursor and Claude Code (setup). Dives draw inline in Claude and ChatGPT (release notes). Other AI apps get a preview and a link (Dives).
  • Price: No separate charge: MCP and Dives are on every plan, including Lite at $0. Business is $250 a month per organisation plus usage (pricing).

MotherDuck has put more work than most vendors here into what comes back. A Dive is a small piece of web code (a React component) plus the SQL behind it, saved in your workspace and run against live data. Inside Claude or ChatGPT it opens as an MCP Apps view, an HTML panel in the chat that you can filter and drill into, and every edit is kept as a version (Dives). Since 12 November a new Dive starts private and you share it with chosen people or the whole organisation (release notes).

Decide who needs query_rw before rolling it out. MotherDuck has no setting for which tools each person gets, so you control that in each AI app’s settings or by handing out read-scaling tokens (read-only guide). On Claude Team and Enterprise, owners can block a tool for everyone (Claude help). For plain DuckDB files there’s a separate local server, read-only unless you start it with —read-write (GitHub).

6. PostHog MCP server

PostHog’s server answers product-usage questions from the events PostHog already collects, for anyone on your team with a PostHog login.

  • Sign-in: Each person, with OAuth, or a personal API key scoped to one project (FAQ).
  • Can it change data: Yes, by default. Adding ?readonly=true to the server URL removes every tool that creates, updates or deletes data (FAQ).
  • Works in: Any AI app that supports hosted MCP servers. PostHog documents Claude, Cursor and others (docs) and lists a ChatGPT plugin (PostHog MCP).
  • Price: Connecting and calling tools is free. The PostHog usage a query consumes is billed as normal (PostHog MCP page).

For product-usage questions we’d start here. It has tools for governed metrics, so a metric like MRR can have one definition whoever asks (tools). The one thing we’d change before a wider team uses it is the read-write default. Charts render in some AI apps. In ChatGPT they may not show yet, as #92709 is still open.

7. Supabase MCP server

Supabase’s server is built for developers working on a Supabase project. Alongside SQL it applies migrations and creates branches and new projects.

  • Sign-in: Each developer, through their Supabase account, choosing an organisation (docs).
  • Can it change data: Yes, unless you add read_only=true to the URL, which runs every query as a read-only Postgres user (docs).
  • Works in: Any AI app that supports hosted MCP servers. Supabase lists ChatGPT and Claude as connectors, plus Claude Code and Cursor (docs).
  • Price: No price for the server itself. Creating projects or branches can add charges, and branching needs a paid plan (docs).

Supabase’s own docs advise against giving it to your customers (docs), and we wouldn’t hand it to a sales team either.

Open-source servers you install yourself

These are open-source servers you run next to Claude Code or Cursor, or host yourself. Most give everyone one shared database login, so everyone gets the same access. ChatGPT on the web only accepts servers that sign each person in with OAuth or need no sign-in, so one that checks a shared token won’t work there. Both points make them a fit for a developer working alone.

8. Postgres MCP server: Postgres MCP Pro

Postgres MCP Pro, a free, MIT-licensed project from Crystal DBA, is built for whoever looks after a Postgres database: explain plans, index tuning and health checks, with SQL alongside.

  • Sign-in: None. Everyone using one instance shares the database role in its connection string (README).
  • Can it change data: Yes, by default. Starting it with the option —access-mode=restricted runs SQL in read-only transactions with a 30-second timeout (source).
  • Works in: AI apps on your own computer: Claude Desktop, Cursor and others. ChatGPT isn’t documented (README).

Of the open-source Postgres servers, this is the one we’d look at first. It had about 620,000 PyPI downloads in the month to 16 November (PyPI stats), and it’s good at the DBA side. For analysts asking business questions it’s thinner, since all it knows is the raw schema. The last release is from May 2025 (PyPI), and later code changes aren’t in a release yet (releases).

9. MySQL MCP server: mcp-server-mysql

mcp-server-mysql is a free, MIT-licensed community server for querying MySQL from Claude Code or Claude Desktop on your own machine.

  • Sign-in: None per person. It uses one MySQL user from environment variables, and when run as a shared server (remote mode) it checks one token, a fixed secret everyone sends (configuration).
  • Can it change data: Not by default. Inserts, updates, deletes and schema changes each need their own setting turned on when you start it (configuration).
  • Works in: AI apps on your own computer: Claude Code, Claude Desktop and Cursor. ChatGPT on the web can’t send the fixed token that remote mode expects (OpenAI auth docs).

Putting every kind of write behind its own switch is the right default, and we wish more servers here did it. In practice it’s a tool for one person running it locally, and about 77,000 npm downloads in the month to 10 November suggest plenty of people do exactly that (npm).

10. SQL Server MCP: Microsoft’s SQL MCP Server

Microsoft’s SQL MCP Server is part of Data API builder, its free, open-source tool for serving database tables through REST and GraphQL APIs. It’s built for teams who want to publish specific tables to an agent instead of opening up SQL.

  • Sign-in: Microsoft Entra ID or another OAuth provider, or no sign-in at all (anonymous). No sign-in is the default in version 2.0, so anyone who can reach the server can use it until you change that (authentication docs).
  • Can it change data: Yes. All seven tools, create, update and delete included, are on by default, and roles you define in its config file decide what each caller may run (tool reference).
  • Works in: Documented for VS Code and Microsoft Foundry. Microsoft says it works with any MCP client, but has no guide for Claude or ChatGPT (overview).

Microsoft went the other way from most servers here. There’s no free-form SQL at all: you list the tables and views to expose (entities) in a config file, roles you define there decide which fields and rows each caller sees, and the agent reads your descriptions. Reads cover one table or view with no joins, so a question across tables means writing a view first (tool reference). We’d switch sign-in to Entra ID before it goes anywhere near a network, given that anonymous is the default.

11. Redshift MCP server (AWS Labs)

AWS Labs’ Redshift server, free under Apache 2.0, is for an engineer querying Redshift from a code editor on their own machine.

  • Sign-in: Your local AWS credentials. Queries run as the Redshift user linked to your AWS login (your IAM identity) (README).
  • Can it change data: Not by default. Each statement runs in a read-only transaction that is then rolled back, and starting it with the setting ACCESS_MODE=read-write turns writes on (docs).
  • Works in: Only AI apps on your own computer: Kiro (Amazon’s AI code editor), Cursor and VS Code (docs). There’s no hosted mode, so ChatGPT on the web needs OpenAI’s Secure MCP Tunnel to reach it (OpenAI help).

It does the safe thing by default. The catch is who can use it: anyone without AWS credentials set up in a terminal, which is most of a business team, can’t.

12. Google MCP Toolbox for Databases

Google’s MCP Toolbox for Databases is a free, open-source server that ships prebuilt tools for each database and lets you publish queries you’ve written and reviewed as named tools, across Postgres, MySQL, SQL Server, BigQuery, Snowflake and many more.

  • Sign-in: Database credentials in its config, with IAM for Google Cloud sources. You can also make it check a sign-in token from your company’s identity provider (OIDC) before it answers (auth docs).
  • Can it change data: It depends on the tools you publish. Free-form execute_sql tools exist for many sources, and Google says they’re for hands-on developer work, not production agents (execute_sql). Read-only mode only covers Google-hosted databases (Cloud SQL for Postgres and MySQL, AlloyDB and BigQuery) (read-only mode). For any other database, we’d use a read-only database user.
  • Works in: AI apps on your own computer: Gemini CLI, Claude Code, Claude Desktop and others. ChatGPT isn’t documented (docs).

Google recommends predefined tools for production agents (introduction): an agent that can only run queries someone wrote and reviewed can’t run one nobody expected. We think that’s right for anything customer-facing. The cost is that the agent can only answer questions someone thought of in advance.

We took the opposite approach in our own server. The agent writes its own SELECT, guided by the descriptions and metrics your team publishes, so it can answer a question nobody wrote a tool for. The safety then comes from refusing anything but a single SELECT (on SQL Server, from a read-only login), which protects your data but can’t promise the query is the one you’d have written.

13. DBHub

DBHub, from Bytebase, is a small, free server, MIT-licensed, for working across Postgres, MySQL, MariaDB, SQL Server, Oracle and SQLite from one place.

  • Sign-in: None per person. When you share it over a network, there’s no sign-in by default. You can add a shared bearer token, one fixed secret every user sends (command line).
  • Can it change data: Yes, by default. With readonly set, the database itself enforces it for Postgres, MySQL, MariaDB and SQLite. On SQL Server and Oracle, use a read-only user as well (execute_sql).
  • Works in: AI apps on your own computer: Claude Desktop, Claude Code, Cursor and others (README).

It’s more widely used than its profile suggests: more than 3,800 GitHub stars (GitHub) and about 166,000 npm downloads in the month to 10 November, against the MySQL server’s 77,000 (npm). It’s handy when you hop between databases in one session. We’d set readonly and keep it off the network, because it accepts any connection with no sign-in until you add a token (command line).

BI tools and semantic layers

If your metrics already live in a BI tool or a semantic layer (one place where metrics and joins are defined once), let the agent read them there. Its numbers then match the dashboards people already trust.

14. Metabase MCP server

Metabase’s MCP server lets Claude or ChatGPT query through what Metabase calls its semantic layer: your metrics, models, segments, saved questions and glossary (tools).

  • Sign-in: Each person, with OAuth from Metabase itself. On Pro and Enterprise it can also run questions on behalf of people signed in to your own product, such as your customers (docs).
  • Can it change data: Yes. execute_sql can overwrite or delete data (tools). We’d connect a read-only database user to stop that, and on Pro and Enterprise admins can block raw SQL per group (usage controls).
  • Works in: Any AI app that supports hosted MCP servers. Metabase documents Claude, ChatGPT, Cursor and VS Code, and draws charts inline once an admin switches each AI app on (docs).
  • Price: On every plan, open source included. Starter is $100 a month for five users and Pro $575 a month for ten (pricing).

If your team already works in Metabase, this is our pick in this group: the agent reads the same metrics, saved questions and glossary your team already built (MCP tools).

15. Looker MCP server

Looker’s managed server queries through your LookML model, which holds the measures and joins you’ve already defined in Looker.

  • Sign-in: Each person, with OAuth, once a Looker admin has enabled each AI app. Claude, ChatGPT and other listed apps take a switch, and the rest need manual registration (docs).
  • Can it change data: Every tool starts disabled, and an admin switches each one on. The admin panel marks which tools read and which write (admin docs).
  • Works in: AI apps that support hosted MCP servers, once a Looker admin has enabled each one. Google documents Claude, ChatGPT, Cursor and others. Your Looker must be one that Looker hosts for you, so a self-hosted Looker can’t use it (docs).
  • Price: No extra charge, but calls count against your instance’s API allowance: 1,000 query calls a month on Standard, 100,000 on Enterprise. Looker itself is sold by quote (pricing).

It came out of preview on 26 October (release notes). The MCP docs warn of occasional timeouts at busy times, because the server runs at a fixed capacity. An admin has to switch on every tool and every AI app, which also means nothing reaches people by accident. On Standard, 1,000 query calls a month is about 50 a working day across the whole team, so check the allowance before rolling it out widely.

16. Power BI: Fabric IQ MCP

Fabric IQ MCP lets an agent query your Power BI semantic models, with each person signed in through Microsoft Entra ID.

  • Sign-in: Each person, through Microsoft Entra ID (docs).
  • Can it change data: No. It’s read-only and keeps Power BI’s row-level security. Editing models needs Microsoft’s separate Authoring server (overview).
  • Works in: Microsoft says any compatible MCP client can connect, but its guides cover its own tools: GitHub Copilot CLI, Microsoft 365 Copilot, Copilot Studio and Microsoft Foundry. Other AI apps need an app registration in your Microsoft Entra tenant (docs, overview).
  • Price: Microsoft lists no price or plan for this server. Power BI Pro is $14 per user a month, paid yearly (pricing).

Microsoft’s guide for Claude Desktop and ChatGPT covers its other Power BI MCP servers, not this one (guide). If your company already runs on Microsoft 365, the Copilot route is the one Microsoft supports.

17. Hex MCP server

Hex’s server hands your question to Hex’s own agent, which works it through in a Hex thread and sends a carousel of charts back to Claude.

  • Sign-in: Each person, with OAuth, as their Hex user (docs).
  • Can it change data: Inside Hex it can create projects, edit notebook cells and run notebooks. Edits only change a project’s draft, never the published version. Hex doesn’t say whether it can write to your database (docs).
  • Works in: Any AI app that supports hosted MCP servers. Hex documents Claude, ChatGPT, Cursor and others (docs).
  • Price: Team and Enterprise plans only. Team is $75 per editor a month (pricing). Using the server needs the Explorer role or above, and editing projects needs Editor (docs). Agent runs use Hex credits on top: Team grants 40 a month per editor and 10 per explorer, a simple count question uses about 1, and automatic top-ups add 50 credits for $25 (credits docs).

Each question typically takes several minutes to answer, and the chart widget expires after 24 hours (docs). If your analysts already work in Hex, the wait buys a full Hex thread behind every answer. For a quick number in a chat, it’s a long time.

18. Cube MCP server

Cube’s server puts Cube’s semantic layer into Claude and ChatGPT, so the agent asks for defined measures instead of writing SQL.

  • Sign-in: Each person, with OAuth, or API keys for your own software. With token exchange, an admin’s API key can act as a specific user, which lets your own product query Cube on each person’s behalf (docs).
  • Can it change data: Its 30 tools can edit the files that define your Cube metrics, commit those changes, publish dashboards and delete reports, depending on each person’s role. A Viewer sees none of the editing tools (docs).
  • Works in: Any AI app that supports hosted MCP servers. Cube documents Claude, ChatGPT and Cursor (docs).
  • Price: On every plan, Free included, which allows 1,000 queries a day. Starter is $40 per developer a month. Viewer seats start on Premium, at $20 per user a month, and Premium is $80 per developer (pricing).

Row-level security plus the ability to act as another user is what you need when the people asking are your own customers, and of the servers here it’s the one we’d look at first for that. The editing tools give an agent more reach than most teams want. Reports and dashboards are saved in Cube, and some AI apps can show interactive views (docs).

19. dbt MCP server

dbt’s server gives an agent your dbt project: your models and how they feed into each other (lineage), plus Semantic Layer metrics on the Starter or Enterprise plan.

  • Sign-in: Each person with OAuth on the hosted server (public beta, Starter and above), or personal access tokens and service tokens (setup).
  • Can it change data: Yes. Its execute_sql doesn’t block write statements, so use a read-only database user for it (tool list).
  • Works in: Hosted or on your own machine. dbt documents Claude and Cursor (docs) and ChatGPT (setup).
  • Price: The hosted server is on every dbt platform plan. The Semantic Layer tools need Starter, at $100 per user a month, or above (docs, pricing).

If your metrics are defined in dbt, this is where an agent should read them.

AI data tools with their own app

These products are built for asking questions of your data with AI. Each has its own app alongside the MCP server, where the charts or notebooks live.

20. Basedash

Basedash is a BI tool with its own agent. Its MCP server hands your question to that agent and builds the charts and dashboards in Basedash.

  • Sign-in: Each person, with a Basedash account over OAuth (docs).
  • Can it change data: Its queries are read-only by default. A change to your data can only be proposed inside the Basedash web app, with your approval, on Postgres, Redshift or Supabase (accuracy and trust). It also works with a read-only database user (data sources).
  • Works in: Any AI app that supports hosted MCP servers. Basedash documents Claude Code, Cursor and ChatGPT (docs).
  • Price: Startup plan, $1,000 a month for up to 25 users with $1,000 of AI credits included, plus usage fees on your next bill if you go over, or Enterprise (pricing).

It makes most sense if you want Basedash as your BI tool anyway and like the idea of asking it from Claude too. As a way to get one database into Claude, it’s an expensive route.

21. Querio

Querio is a notebook tool you drive from the chat: the agent creates notebooks, runs Python and publishes changes to your Querio workspace.

  • Sign-in: Each person, in the browser, over OAuth (reference).
  • Can it change data: It creates notebooks, runs Python, publishes changes to your workspace and can open pull requests in GitHub or Bitbucket (reference). Querio advises connecting databases with a read-only identity (data docs).
  • Works in: Any AI app that supports hosted MCP servers. Querio documents ChatGPT and Claude (reference).
  • Price: MCP is on every plan, Free included. A Querio credit pays for one AI model call in Querio, MCP calls don’t use credits, and extra credits cost $0.50 each on Free and $0.25 on Core (pricing).

Core is $2,499 a month, or $25,000 paid annually (pricing). Notebooks are the whole idea here, so if what you want is a quick answer to a one-off question, it’s built for something else.

22. BlazeSQL

BlazeSQL is the closest thing here to what we built: read-only SQL with interactive graphs in the chat, across seven databases, including SQL Server, Postgres, MySQL and Redshift (connectors).

  • Sign-in: Each person, with a BlazeSQL account or Google sign-in over OAuth (docs).
  • Can it change data: No. execute_sql runs read-only queries. BlazeSQL still recommends a read-only database login, since access depends on the credentials you give it (docs).
  • Works in: Any AI app that supports hosted MCP servers. BlazeSQL documents Claude, ChatGPT, Cursor and others (docs).
  • Price: The MCP server needs Team Advanced (docs), which is $800 a month for three users. The packaged Claude and ChatGPT connectors are listed on Team, at $400 a month for three users (pricing).

Check which of the two you need before you budget. Databases you connected only in its desktop app also need that app open for Claude to reach them (Claude connector).

Should you build your own MCP server?

One data team we spoke to built their own, and it works well, though it took their analytics engineer three months with no other projects. Their agent runs inside Claude Desktop as a plugin and mostly plans: it reads a business context file of 5,000 to 6,000 lines and picks one of their semantic views. Claude then runs the SQL through the Snowflake MCP server.

For harder questions their agent traces a number back through the dbt lineage and checks its answer against the official KPI definition. The failure they designed against was an agent adding a filter to marketing cost halfway through and returning a number no dashboard in the company shows. The plugin has 13 skills, and a stop hook sends each session to their tracing tool. What people typed is masked for anyone who declines to share it. Around 300 people use it, and it scores 92% on a set of hard questions they wrote themselves.

We’d only suggest copying them if you’re in a similar spot. They started with more than 1,000 dbt models already written and could spare an engineer for a full quarter, and most of the teams we talk to are missing one or both of those, in which case one of the servers above is the quicker start.

Frequently asked questions

What is a database MCP server?

A small service that gives an AI assistant tools to read a database through the Model Context Protocol, which Anthropic published in November 2024. The assistant calls a tool, the server runs the query, and the rows come back into the conversation.

Is it safe to connect Claude to a production database?

Only when neither the server nor the database user can write. Use a read-only user, switch on the server’s read-only mode where it has one, and point it at a replica if load worries you. On Claude Team and Enterprise, owners can also block a server’s write tools for everyone.

Can ChatGPT use an MCP server?

Yes. On the website, full support is on Business, Enterprise and Edu, where admins add servers, and Pro accounts get read and fetch tools only (OpenAI help). The server has to be reachable over HTTPS and sign people in with OAuth or not at all. A server that only runs on your own computer needs OpenAI’s Secure MCP Tunnel or the ChatGPT desktop app (OpenAI docs).

What’s the difference between an MCP server and uploading a CSV to Claude?

A file is a snapshot that has to fit in the conversation. An MCP server queries the live database, so the answer reflects today’s data, and a server that signs each person in also applies their own permissions, so nobody sees rows they shouldn’t. Servers that share one login give everyone the same access.

Do I need a semantic layer?

Not to start. What an agent needs first is descriptions: what each table and column means, how tables join, and how your key metrics are defined. If you already keep a semantic layer in Snowflake, Looker, Cube, dbt or Metabase, use the server that reads it. In Bonnard, your team drafts metrics and descriptions per data source in the Context section of our console.


Every fact here was checked against the vendor’s own pages, linked in each entry, on 16 November 2026. If something has changed, tell us through bonnard.ai/get-in-touch and we’ll fix it.

Get started

Put BI where the work happens

Connect a database, create an app, send a link.

No seats · No minimums · No card to start

Keep reading