SQL Sentinel MCP Server

io.github.tkmawarire由 GitHub 账号 tkmawarire 发布v2.3.1更新于 Oct 24, 2026

SQL Server monitoring and diagnostics for AI agents using Extended Events. No ODBC drivers required.

已验证STDIO仅桌面数据库安全与监控

概览

AI 生成的概览

一个本地 MCP 服务器,让 AI 助手通过扩展事件会话、等待统计和直接查询来监控、诊断和操作 SQL Server。

可以做什么
创建、启动、停止、删除和列出扩展事件捕获会话,包括一步快速捕获和短时实时事件流,支持过滤、排序和去重,以及查询指纹归一化和执行顺序分析。按指纹、数据库、应用或登录获取聚合统计,并查看数据库、应用、登录、会话和阻塞情况。诊断死锁、被阻塞进程和等待统计,并运行涵盖慢查询、死锁、阻塞和洞察的综合健康检查。检查或授予所需的 SQL Server 权限。执行数据库操作:列出表、描述表结构、执行 SELECT 查询、插入、更新、创建和删除表。
适用场景
适合需要助手排查 SQL Server 性能问题(如慢查询、N+1 模式、死锁或阻塞)的场景,也适合在你管理的 SQL Server 实例上做常规表结构查看和数据查询。
运行要求
需要本地运行环境:Docker、用于 NuGet 全局工具的 .NET 9 SDK,或从源码构建。需要可访问的 SQL Server 2012 及以上实例并启用扩展事件。连接字符串须通过环境变量 SQL_SENTINEL_CONNECTION_STRING 提供。登录账号需要 ALTER ANY EVENT SESSION 和 VIEW SERVER STATE 权限;阻塞进程检测还需配置 blocked process threshold。
安装前请注意
环境变量 SQL_SENTINEL_CONNECTION_STRING 含数据库凭据,须妥善保管。工具可创建、插入、更新和删除表以及授予权限,因此可能产生写入和破坏性变更。事件中捕获的查询文本可能含敏感数据。不要在生产环境长期保留捕获会话,并只授予最低必要权限。

安装

在 SourceWeft 中

  1. 打开 控制台中的 SQL Sentinel MCP Server,将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

SQL Sentinel MCP Server

[NuGet] [Docker] [License: MIT]

A production-ready MCP (Model Context Protocol) server for SQL Server monitoring, diagnostics, and database operations. Built with .NET 9 and Microsoft.Data.SqlClient for native SQL Server connectivity — no ODBC drivers required.

Features

  • Session Management — Create, start, stop, drop, and list Extended Events sessions
  • Smart Filtering — Filter by application, database, user, duration, host, and text patterns
  • Query Fingerprinting — Normalize and group similar queries differing only in literal values
  • Sequence Analysis — Trace execution order with timing gaps and cumulative duration
  • Deadlock Detection — Capture and analyze XML deadlock reports with victim/process details
  • Blocking Analysis — Monitor blocked process events with wait resource and SQL text
  • Wait Stats — Query sys.dm_os_wait_stats directly, categorized by type (CPU, I/O, Lock, Memory, etc.)
  • Health Check — Comprehensive server diagnostic: slow queries, deadlocks, blocking, wait stats, and insights
  • Real-Time Streaming — Stream captured events for a specified duration
  • Production-Safe — Auto-excludes noise (sp_reset_connection, SET statements, trace queries)
  • Database Operations — List tables, describe schemas, query data, insert, update, and drop tables
  • AI-Optimized — Structured JSON output with optional Markdown formatting

Requirements

  • SQL Server 2012+ with Extended Events enabled (default)
  • Required permissions:
    sql
    GRANT ALTER ANY EVENT SESSION TO [your_login];GRANT VIEW SERVER STATE TO [your_login];
  • For blocked process detection:
    sql
    EXEC sp_configure 'show advanced options', 1;RECONFIGURE;EXEC sp_configure 'blocked process threshold', 5;RECONFIGURE;

Installation

Option 1: Docker (Recommended)

No .NET SDK required. Works on any system with Docker installed.

bash
docker pull ghcr.io/tkmawarire/sql-sentinel-mcp:latest
Claude Desktop (claude_desktop_config.json)
json
{  "mcpServers": {    "sql-sentinel": {      "command": "docker",      "args": ["run", "-i", "--rm", "--network", "host",               "-e", "SQL_SENTINEL_CONNECTION_STRING=Server=localhost;Database=master;User Id=sa;Password=YourPassword;TrustServerCertificate=true",               "ghcr.io/tkmawarire/sql-sentinel-mcp:latest"]    }  }}
Claude Code
bash
claude mcp add sql-sentinel \  -e SQL_SENTINEL_CONNECTION_STRING="Server=localhost;Database=master;User Id=sa;Password=YourPassword;TrustServerCertificate=true" \  -- docker run -i --rm --network host \  -e SQL_SENTINEL_CONNECTION_STRING \  ghcr.io/tkmawarire/sql-sentinel-mcp:latest

Network access: The -i flag is required for stdio transport. Use --network host so the container can reach SQL Server on your host machine. For remote SQL Server, omit --network host and use the accessible hostname in your connection string.

Connection string: Set SQL_SENTINEL_CONNECTION_STRING via -e. All tools read the connection string from this environment variable.

Option 2: .NET Global Tool (NuGet)

Requires .NET 9 SDK or later.

bash
dotnet tool install -g SQLSentinel.Mcp
json
{  "mcpServers": {    "sql-sentinel": {      "command": "sql-sentinel-mcp",      "env": {        "SQL_SENTINEL_CONNECTION_STRING": "Server=localhost;Database=master;User Id=sa;Password=YourPassword;TrustServerCertificate=true"      }    }  }}

Option 3: Build from Source

bash
git clone https://github.com/tkmawarire/sql-sentinel.gitcd sql-sentineldotnet build

Run directly:

bash
dotnet run --project SqlServer.Profiler.Mcp/

Or publish a self-contained single binary:

bash
# Windowsdotnet publish SqlServer.Profiler.Mcp/ -c Release -r win-x64 --self-contained
# Linuxdotnet publish SqlServer.Profiler.Mcp/ -c Release -r linux-x64 --self-contained
# macOS (Apple Silicon)dotnet publish SqlServer.Profiler.Mcp/ -c Release -r osx-arm64 --self-contained
# macOS (Intel)dotnet publish SqlServer.Profiler.Mcp/ -c Release -r osx-x64 --self-contained

Output will be in bin/Release/net9.0/{runtime}/publish/

Connection Strings

All tools read the connection string from the SQL_SENTINEL_CONNECTION_STRING environment variable. Set it once before starting the server:

bash
export SQL_SENTINEL_CONNECTION_STRING="Server=localhost;Database=master;User Id=sa;Password=YourPassword;TrustServerCertificate=false;Encrypt=true"

SQL Authentication:

Server=localhost;Database=master;User Id=sa;Password=YourPassword;TrustServerCertificate=false;Encrypt=true

Windows Authentication:

Server=localhost;Database=master;Integrated Security=true;TrustServerCertificate=false;Encrypt=true

Note: Only use TrustServerCertificate=true in development environments with self-signed certificates. For production, always use TrustServerCertificate=false with a valid SSL certificate.

Azure SQL:

Server=yourserver.database.windows.net;Database=yourdb;User Id=user;Password=password;Encrypt=true

MCP Tools Reference

Session Lifecycle

ToolDescription
sqlsentinel_create_sessionCreate an Extended Events session with filters (not started)
sqlsentinel_start_sessionStart capturing events for an existing session
sqlsentinel_stop_sessionStop capturing; events are retained
sqlsentinel_drop_sessionDrop session and discard all events
sqlsentinel_list_sessionsList all MCP-created sessions with state and buffer usage
sqlsentinel_quick_captureCreate and start a session in one step

Event Retrieval

ToolDescription
sqlsentinel_get_eventsRetrieve captured events with filtering, sorting, and deduplication
sqlsentinel_get_statsAggregate statistics grouped by fingerprint, database, app, or login
sqlsentinel_analyze_sequenceAnalyze query execution sequence with timing and gaps
sqlsentinel_get_connection_infoList databases, applications, logins, sessions, and blocking info
sqlsentinel_stream_eventsReal-time event capture for a specified duration (1–300s)

Diagnostics

ToolDescription
sqlsentinel_get_deadlocksRetrieve deadlock events with victim, processes, locks, and SQL text
sqlsentinel_get_blockingRetrieve blocked process events with wait resources and SQL text
sqlsentinel_get_wait_statsQuery sys.dm_os_wait_stats categorized by type (no session required)
sqlsentinel_health_checkComprehensive report: slow queries, deadlocks, blocking, wait stats, insights

Permissions

ToolDescription
sqlsentinel_check_permissionsCheck current login permissions and blocked process threshold config
sqlsentinel_grant_permissionsGrant required permissions to a login (requires sysadmin)

Database Operations

ToolDescription
sqlsentinel_list_tablesList all user tables in the database (schema-qualified)
sqlsentinel_describe_tableDetailed table schema: columns, indexes, constraints, foreign keys
sqlsentinel_create_tableCreate a new table via CREATE TABLE statement
sqlsentinel_insert_dataInsert data via INSERT statement
sqlsentinel_read_dataExecute SELECT queries and return results
sqlsentinel_update_dataUpdate data via UPDATE statement
sqlsentinel_drop_tableDrop a table via DROP TABLE statement

Usage Examples

Quick Debug Session

Agent: sqlsentinel_quick_capture(    sessionName: "debug_api",    applications: "MyWebApp",    minDurationMs: 100)
// User triggers the slow operation
Agent: sqlsentinel_get_events(    sessionName: "debug_api",    sortBy: "DurationDesc",    limit: 20)
Agent: sqlsentinel_drop_session(sessionName: "debug_api")

Find N+1 Queries

Agent: sqlsentinel_quick_capture(    sessionName: "n_plus_one_check",    databases: "OrdersDB")
// User loads a page
Agent: sqlsentinel_get_stats(    sessionName: "n_plus_one_check",    groupBy: "QueryFingerprint")
// Look for queries with high execution counts

Trace Specific Operation

Agent: sqlsentinel_analyze_sequence(    sessionName: "my_session",    correlationId: "order-12345",    responseFormat: "Markdown")

Deadlock Detection

Agent: sqlsentinel_quick_capture(    sessionName: "deadlock_monitor",    eventTypes: "Deadlock")
// Wait for deadlocks to occur
Agent: sqlsentinel_get_deadlocks(    sessionName: "deadlock_monitor",    responseFormat: "Markdown")

Blocking Analysis

Agent: sqlsentinel_quick_capture(    sessionName: "blocking_check",    eventTypes: "BlockedProcess")
// Requires: sp_configure 'blocked process threshold', 5
Agent: sqlsentinel_get_blocking(    sessionName: "blocking_check",    responseFormat: "Markdown")

Server Health Check

Agent: sqlsentinel_health_check(    sessionName: "my_session",    slowQueryThresholdMs: 1000,    responseFormat: "Markdown")

Database Operations

Agent: sqlsentinel_list_tables()
Agent: sqlsentinel_describe_table(    name: "dbo.Products")
Agent: sqlsentinel_read_data(    sql: "SELECT TOP 10 * FROM dbo.Products ORDER BY CreatedDate DESC")

Wait Stats (No Session Required)

Agent: sqlsentinel_get_wait_stats(    topN: 20,    responseFormat: "Markdown")

Query Fingerprinting

Queries are normalized to group similar ones:

sql
-- These become one fingerprint:SELECT * FROM Users WHERE id = 123SELECT * FROM Users WHERE id = 456
-- Fingerprint: abc123:SELECT * FROM Users WHERE id = ?-- Execution count: 2

Noise Filtering

Default excluded patterns (when excludeNoise=true):

  • sp_reset_connection — Connection pool reset
  • SET TRANSACTION ISOLATION LEVEL — Session setup
  • SET NOCOUNT, SET ANSI_* — Client configuration
  • sp_trace_*, fn_trace_* — Trace system queries

Supported Event Types

SqlBatchCompleted, RpcCompleted, SqlStatementCompleted, SpStatementCompleted, Attention, ErrorReported, Deadlock, BlockedProcess, LoginEvent, SchemaChange, Recompile, AutoStats

Project Structure

sql-profiler-mcp/├── .github/│   └── workflows/│       ├── docker.yml                     # Build & push multi-arch Docker images│       └── publish-mcp-registry.yml       # Publish NuGet + MCP registry├── .mcp/│   └── server.json                        # MCP manifest (NuGet + OCI packages)├── SqlServer.Profiler.Mcp/                # Main MCP server (stdio transport)│   ├── SqlServer.Profiler.Mcp.csproj│   ├── Program.cs                         # Entry point, DI setup, MCP config│   ├── Models/│   │   ├── ProfilerModels.cs              # Records, enums, data models│   │   └── DbOperationResult.cs           # Result model for CRUD operations│   ├── Services/│   │   ├── ProfilerService.cs             # Core Extended Events logic│   │   ├── QueryFingerprintService.cs     # SQL normalization & fingerprinting│   │   ├── WaitStatsService.cs            # DMV-based wait stats analysis│   │   ├── SessionConfigStore.cs          # In-memory session config storage│   │   └── EventStreamingService.cs       # Real-time event streaming│   ├── Utilities/│   │   └── SqlInputValidator.cs           # SQL input validation & escaping│   └── Tools/│       ├── SessionManagementTools.cs      # Session lifecycle tools (6)│       ├── EventRetrievalTools.cs         # Event retrieval tools (5)│       ├── DiagnosticTools.cs             # Diagnostic tools (4)│       ├── PermissionTools.cs             # Permission tools (2)│       └── DatabaseTools.cs               # Database CRUD tools (7)├── SqlServer.Profiler.Mcp.Api/            # Debug REST API (Swagger on port 5100)│   ├── SqlServer.Profiler.Mcp.Api.csproj│   ├── Program.cs│   ├── Controllers/│   │   └── ProfilerController.cs│   ├── Models/│   │   └── RequestModels.cs│   └── appsettings.json├── SqlServer.Profiler.Mcp.Cli/            # Debug CLI (REPL + script mode)│   ├── SqlServer.Profiler.Mcp.Cli.csproj│   └── Program.cs├── SqlServer.Profiler.Mcp.Tests/          # xUnit tests for core MCP library (228 tests)│   └── ...├── SqlServer.Profiler.Mcp.Api.Tests/      # xUnit tests for API project (29 tests)│   └── ...├── Dockerfile                             # Multi-stage build (bookworm-slim)├── .dockerignore├── SqlServer.Profiler.Mcp.slnx           # Solution file├── CLAUDE.md├── CONTRIBUTING.md└── README.md

Development

Prerequisites

  • .NET 9 SDK
  • SQL Server 2012+ instance (local, Docker, or remote)
  • Docker (optional, for container builds)

Clone & Build

bash
git clone https://github.com/tkmawarire/sql-sentinel.gitcd sql-sentineldotnet restoredotnet build

Running the MCP Server Locally

bash
dotnet run --project SqlServer.Profiler.Mcp/

The server communicates over stdio using the MCP protocol. Connect it to an MCP client (Claude Desktop, Claude Code, etc.) for interactive use.

Using the Debug API

The API project provides a REST wrapper around all MCP tools with Swagger UI for manual testing.

bash
dotnet run --project SqlServer.Profiler.Mcp.Api/
  • Swagger UI: http://localhost:5100/
  • Configure the connection string via environment variable SQL_SENTINEL_CONNECTION_STRING

Using the Debug CLI

The CLI project provides an interactive REPL and script mode for testing tools directly.

bash
# Interactive REPL modedotnet run --project SqlServer.Profiler.Mcp.Cli/
# List all available toolsdotnet run --project SqlServer.Profiler.Mcp.Cli/ list
# Get help for a specific tooldotnet run --project SqlServer.Profiler.Mcp.Cli/ help sqlsentinel_quick_capture
# Execute a single tooldotnet run --project SqlServer.Profiler.Mcp.Cli/ call sqlsentinel_list_sessions

Set the SQL_SENTINEL_CONNECTION_STRING environment variable before running.

Docker Build

bash
docker build -t sql-sentinel-mcp:test .docker run -i --rm --network host sql-sentinel-mcp:test

Architecture

Key Patterns

  • Dependency injection via Microsoft.Extensions.Hosting
  • stdio transport — stdout is reserved for MCP protocol; all logging goes to stderr
  • Tool auto-discovery — MCP tools are discovered from the assembly via WithToolsFromAssembly()
  • XE session prefix — All created sessions are prefixed with mcp_sentinel_
  • Two event shapes — Standard events (query, login, recompile) with typed fields, and XML-payload events (deadlock, blocking) parsed from Extended Events XML

Adding a New MCP Tool

  1. Create a public static method in the appropriate file under Tools/ (or create a new file)
  2. Decorate with [McpServerTool(Name = "sqlsentinel_your_tool")] and [Description("...")]
  3. Add parameters with [Description("...")] attributes — they become the tool's input schema
  4. Inject services via method parameters (e.g., IProfilerService, IWaitStatsService)
  5. Return a string (JSON or Markdown) — the framework handles MCP response wrapping
csharp
[McpServerTool(Name = "sqlsentinel_example")][Description("Description shown to AI agents")]public static async Task<string> Example(    IProfilerService profilerService,    [Description("Optional filter")] string? filter = null){    var connectionString = ConnectionStringResolver.Resolve();    // Implementation    return JsonSerializer.Serialize(result);}

Troubleshooting

"Permission denied" creating session

sql
GRANT ALTER ANY EVENT SESSION TO [your_login];GRANT VIEW SERVER STATE TO [your_login];

"Login failed"

  • Check connection string credentials
  • For Windows auth, ensure process runs under correct user
  • For Azure SQL, ensure firewall allows your IP

No events captured

  1. Verify session is RUNNING (sqlsentinel_list_sessions)
  2. Check filters aren't too restrictive
  3. Verify target database/app is generating queries
  4. Check minDurationMs isn't filtering everything

No deadlock events

  • Ensure session was created with eventTypes: "Deadlock"
  • Deadlocks must actually occur while the session is running

No blocking events

  • Ensure blocked process threshold is configured: sp_configure 'blocked process threshold', 5
  • Ensure session was created with eventTypes: "BlockedProcess"
  • Blocking must exceed the configured threshold (seconds)

Timeout reading events

Large ring buffers with many events can be slow to parse. Use:

  • Time filters to narrow the window
  • Increase command timeout in code if needed

Security Notes

  • The SQL_SENTINEL_CONNECTION_STRING environment variable contains credentials — secure appropriately
  • Don't leave sessions running indefinitely on production
  • Query text may contain sensitive data
  • Grant minimum required permissions

Contributing

See CONTRIBUTING.md for guidelines on submitting issues and pull requests.

License

MIT

来源:README.md,提交 84ac172

工具

0
SourceWeft 尚未在沙箱中启动过该软件包,因此其工具未知。

版本历史

1
  1. v2.3.1最新Sep 16, 2026

评分

已满足适用检查项的 50%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分