SQL MCP server

SQL Server MCP for people who manage dozens of instances

One MCP entry, every SQL Server you administer. Connections live in a single connections.json, grouped by client or environment, hot-reloaded without restarting your AI agent — plus execution plans, index audits and stored-procedure analysis.

Built for DBAs and consultants, not for a demo against a single localhost database.

Install

Add this to your AI agent's MCP configuration:

{
  "mcpServers": {
    "sqlserver": {
      "command": "npx",
      "args": ["-y", "@cevelas/mcp-sqlserver"]
    }
  }
}

Then create your connections file and restart the agent:

npx -y @cevelas/mcp-sqlserver --init

That writes ~/.mcp-sqlserver/connections.json from a commented template. Edit it, and ask your agent to "list all SQL Server connections".

Needs Node.js 22 or newer.

Agent	Config file
Claude Desktop (Windows)	%APPDATA%\Claude\claude_desktop_config.json
Claude Desktop (macOS)	~/Library/Application Support/Claude/claude_desktop_config.json
Claude Code	claude mcp add sqlserver -- npx -y @cevelas/mcp-sqlserver
VS Code / Copilot	.vscode/mcp.json
Cursor	~/.cursor/mcp.json

Any MCP-compatible agent works — ChatGPT, Gemini, Copilot, Cline, Zed and others all take the same command / args pair.

Why this one

Most SQL Server MCP servers take a single connection string. That is fine for one database. It falls apart when you administer thirty across eight clients, because every instance needs its own entry in the agent config, its own credentials, and its own restart when something changes.

	Single-DSN servers	This one
Instances per MCP entry	1	all of them
Organised by client or environment	—	connectionGroup
Add or change a connection	edit agent config, restart	edit a file, reload_connections
Which server answered?	you assume	in every response's metadata
Beyond SELECT	—	execution plans, index layout, SP source
Read-only safety rail	—	"readOnly": true per connection

Connections

{
  "connections": [
    {
      "name": "acme-prod",
      "connectionGroup": "Acme Corp",
      "description": "Production - head office",
      "server": "192.168.1.10\\SQLEXPRESS",
      "database": "AcmeDB_Prod",
      "user": "app_reader",
      "password": "${env:ACME_PROD_PASSWORD}",
      "port": 1433,
      "encrypt": true,
      "trustServerCertificate": false,
      "readOnly": true
    }
  ]
}

Field	Required	Notes
name	yes	Unique; this is what you say to the agent
server	yes	Hostname, IP, or host\instance
connectionGroup	no	Client, project or environment. Groups the listing
description	no	Shown in list_connections and in every response
database	no	Defaults to the login's default database
user, password	no	Omit for domain or Entra ID auth
port	no	Defaults to 1433
encrypt, trustServerCertificate	no	encrypt defaults to true
readOnly	no	Rejects writing statements — see below

Anything else you put here is passed straight to mssql, so requestTimeout, connectionTimeout, pool, authentication and a nested options object all work.

Keeping passwords out of the file

Any string may reference an environment variable:

"password": "${env:ACME_PROD_PASSWORD}"

A connection that references a variable you have not set is disabled, and list_connections names both the connection and the missing variable. Leaving the literal in place would only move the failure to connect time, where it arrives as Login failed for user and tells you nothing.

You can also skip the file entirely and pass the whole thing through the agent config, which keeps credentials in one place with the rest of your MCP secrets:

{
  "mcpServers": {
    "sqlserver": {
      "command": "npx",
      "args": ["-y", "@cevelas/mcp-sqlserver"],
      "env": {
        "MSSQL_MCP_CONNECTIONS_JSON": "{\"connections\":[{\"name\":\"prod\",\"server\":\"10.0.0.1\",\"database\":\"App\",\"user\":\"reader\",\"password\":\"...\"}]}"
      }
    }
  }
}

Where the file is looked for

In order, first hit wins:

1. --connections <path>
2. $MSSQL_MCP_CONNECTIONS — a path
3. $MSSQL_MCP_CONNECTIONS_JSON — the JSON itself, inline
4. ./connections.json in the working directory
5. ~/.mcp-sqlserver/connections.json
6. connections.json next to the installed package

A path given explicitly via 1 or 2 that does not exist is an error — the server will not quietly fall back to a different file and talk to the wrong database.

Checking every connection at once

npx -y @cevelas/mcp-sqlserver --check

Connects to every entry in parallel, runs SELECT 1, and lists which ones fail and why. Exits non-zero if any did, so it works in a scheduled task. Useful after a password rotation, or before blaming the agent.

Windows domain and Entra ID authentication

The bundled tedious driver supports NTLM and the Entra ID (Azure AD) family. Add domain for NTLM:

{
  "name": "warehouse",
  "server": "dwh.corp.local",
  "database": "DWH",
  "domain": "CORP",
  "user": "svc_analytics",
  "password": "${env:DWH_PASSWORD}"
}

{
  "name": "azure-sql",
  "server": "myserver.database.windows.net",
  "database": "reporting",
  "encrypt": true,
  "authentication": {
    "type": "azure-active-directory-password",
    "options": { "userName": "${env:AZURE_USER}", "password": "${env:AZURE_PASSWORD}" }
  }
}

Fully integrated auth — a trusted connection with no password at all — needs the native msnodesqlv8 driver, which is not bundled because it would break the one-line install on machines without a build toolchain. NTLM with an explicit service account is the supported path.

Tools

Tool	Arguments	What it does
list_connections	—	Every connection, grouped
reload_connections	—	Re-read the file, drop open pools
query	connection, sql, maxRows?, format?	Run a query
get_schema	connection, table?	Columns, types, nullability, defaults
get_indexes	connection, table	Indexes, types, key and included columns
get_execution_plan	connection, sql	SHOWPLAN_XML — the plan, without running the query
get_stored_procedure	connection, name	Source of a stored procedure

Every response carries the connection it came from:

{
  "metadata": {
    "connection": "acme-prod",
    "connectionGroup": "Acme Corp",
    "description": "Production - head office",
    "server": "192.168.1.10\\SQLEXPRESS",
    "database": "AcmeDB_Prod"
  },
  "data": [ ... ]
}

With thirty connections in play, that line is what tells you the answer came from the client you meant.

Query results that fit in a context window

An agent that runs SELECT * FROM Orders does not need four million rows, and neither does its context window.

• Row cap. query returns at most 1000 rows across all result sets. maxRows changes that, up to 10000. When the cap is hit the query is cancelled on the server, not trimmed afterwards.
• A truncated result says so, first. The response opens with "truncated": true and a note stating that more rows exist, that the total is unknown, and how to get it. It sits before the rows on purpose: a 10000-row response runs to hundreds of KB, and a client that cuts long output keeps the beginning. An agent must never conclude "the table has 10000 rows" from a capped result. Asking for more than 10000 is not an error — the cap is lowered to 10000, and if rows were left out the note says what was asked for.
• format: "csv". JSON repeats every column name on every row. With csv the response is one JSON header line with the metadata, a blank line, then the rows. NULL is an empty cell and an empty string is "", so the two stay distinguishable.
• Server messages. PRINT output and SET STATISTICS IO, TIME ON results come back in messages, so the agent can read logical reads and CPU time instead of guessing.
• Every result set. A batch with several SELECTs returns them all in recordsets. data is still the first one.
• Binary columns come back as hex (0x8f3a...), truncated after 64 bytes, instead of a JSON array with one number per byte.

rowCount is always present; truncated, messages and recordsets only appear when they have something to say. Responses are compact JSON.

What this gets you

Things that are tedious by hand and become one sentence to the agent:

• "Why is this stored procedure slow?" — get_stored_procedure for the source, get_execution_plan for the plan, get_indexes for what is missing.
• "Compare the Orders schema between acme-prod and acme-qa" — get_schema on both, agent diffs them.
• "Which indexes on this table are never covering anything?" — get_indexes plus the queries you care about.
• "I added a client to connections.json" — reload_connections, no restart.

Read-only connections

"readOnly": true

Rejects INSERT, UPDATE, DELETE, MERGE, DROP, TRUNCATE, ALTER, CREATE, GRANT, EXEC, BACKUP, DBCC, OPENQUERY, DISABLE/ENABLE and friends before the query leaves your machine. It also requires the batch to start with something that reads — SELECT, WITH, DECLARE, SET, IF and so on — because T-SQL lets you call a procedure without EXEC, and sp_rename 'dbo.Users','Users_old' contains no blocked keyword at all.

String literals, comments and bracketed identifiers are ignored, so WHERE note = 'please delete this', SELECT [delete] FROM [Audit] and DECLARE @Create DATETIME all pass. get_execution_plan still works, because SHOWPLAN_XML returns the plan without executing anything.

Anything other than an explicit false turns the guard on — a hand-written "readOnly": "false" locks the connection down rather than silently opening it, and says so in list_connections.

This is a guard rail, not a security boundary. It stops an agent from "helpfully" fixing a row in production. It will not stop someone determined to write. The real protection is a SQL login that only has db_datareader:

CREATE LOGIN mcp_reader WITH PASSWORD = '...';
CREATE USER mcp_reader FOR LOGIN mcp_reader;
ALTER ROLE db_datareader ADD MEMBER mcp_reader;
GRANT VIEW DEFINITION TO mcp_reader;   -- for get_stored_procedure
GRANT SHOWPLAN TO mcp_reader;          -- for get_execution_plan

Use both.

Security notes

• connections.json holds credentials. Keep it out of version control — the bundled .gitignore covers connections*.json.
• Prefer ${env:VAR} over literal passwords.
• Give each connection the least privilege it needs. Do not use sa.
• Restrict file permissions: icacls connections.json /inheritance:r /grant:r "%USERNAME%:F" on Windows, chmod 600 connections.json elsewhere.
• The query tool runs whatever SQL the agent writes. That is the point of the tool — treat the connection's permissions as the boundary, not the tool.

Web UI

web/connections.html is a standalone page for editing connections.json without hand-writing JSON: drag and drop between groups, duplicate a connection, autocompleting group selector, and auto-save through the File System Access API in Chrome and Edge. No build, no dependencies, entirely optional. See web/README.md.

Claude Desktop one-click install

Grab the .mcpb bundle from the latest release and drag it onto Claude Desktop's extensions settings. Point it at your connections.json in the extension's settings, or leave that empty to use ~/.mcp-sqlserver/connections.json.

Upgrading from 2.x

Your existing connections.json works unchanged — every new field is optional and the tools take the same arguments. Two things worth knowing:

• Multi-connection was broken before 3.0. The server used the mssql global connection pool, which ignores the config it is handed once a connection is already open. In practice every connection after the first silently reused the first one's server and database. If you were relying on results from more than one connection in a session, they may not have come from where you thought. Fixed in 3.0 with a pool per connection.
• If your agent config points at node C:\path\to\index.js, that still works. The file resolution order now checks that path last, so nothing moves.

Development

npm install
npm test                     # unit tests, no database needed
node .github/scripts/smoke.mjs   # packs, installs and speaks MCP to the tarball

Against a real server, name a connection from your own file:

MSSQL_TEST_CONNECTION=local npm run test:integration

The mssql driver is injected, so the unit tests mock only that boundary — everything else is the real code path. test/contract.test.js freezes the tool names and arguments so a refactor cannot change the MCP surface by accident.

License

MIT — see LICENSE.

Author

Christian Velasquez — @cvelasquez

Issues · Changelog · Sponsor