An MCP server that gives AI assistants safe, read-only access to SQL Server databases. Every query is parsed into a full AST using Microsoft's official T-SQL parser — not regex — so comment injection, string literal tricks, and encoding bypasses are blocked at the syntax level.
MCP server providing AI assistants with safe, read-only SQL Server access
Verify by asking your assistant to call list_servers
Start with Installation for exact commands and file paths.
Why This Approach
AST-level query validation — Most MCP database servers use keyword blocking or no validation at all. This project parses every query into a full syntax tree using Microsoft's official TSql180Parser. Comment injection, string literal tricks, and encoding bypasses are blocked at the syntax level, not with fragile regex patterns.
Rate limiting — Token bucket throughput limiting and concurrency control prevent runaway AI query loops from overwhelming production SQL Servers. No other MCP database server offers this.
DBA diagnostic tooling — Integrated support for First Responder Kit, DarlingData, and sp_WhoIsActive with parameter blocking that prevents write operations. This is an entirely new MCP capability category.
Response size optimisation — DBA tools exclude verbose columns (XML query plans, deadlock graphs, metric breakdowns) and truncate long strings by default, reducing response sizes by 90–99%. Use verbose and includeQueryPlans parameters to get full untruncated output when needed.
Progressive discovery — Up to 31 tools organized into toolsets that load on demand. Only 6 core tools are exposed initially, keeping the AI's context window small and reducing token usage. Additional toolsets are discovered and enabled as needed.
Features
Security
Read-only by design — only SELECT and CTE queries are permitted
AST-based query validation using ScriptDom (not regex)
Parameter blocking on all diagnostic stored procedures to prevent writes
Concurrency and throughput rate limiting
Database Tooling
Multi-server support — named connections to multiple SQL Server instances
Schema overview — concise Markdown schema maps with PKs, FKs, constraints, and defaults
Table documentation — Markdown descriptions of columns, indexes, foreign keys, and constraints
ER diagram generation — PlantUML and Mermaid diagrams with smart cardinality detection
# Linux/macOS
mkdir -p ~/.config/sqlaugur
# Edit ~/.config/sqlaugur/appsettings.json with your server connections
# Windows (PowerShell)
mkdir "$env:APPDATA\sqlaugur" -Force
# Edit %APPDATA%\sqlaugur\appsettings.json with your server connections
Example appsettings.json to save at that location:
# Volume-mount a config file
docker run -i --rm \
-v /path/to/appsettings.json:/app/appsettings.json:ro,Z \
ghcr.io/mbentham/sqlaugur:latest
# Or use environment variables (no config file needed)
docker run -i --rm \
-e SqlAugur__Servers__production__ConnectionString="Server=host.docker.internal;Database=master;..." \
ghcr.io/mbentham/sqlaugur:latest
Note: To reach a SQL Server on the host machine, use host.docker.internal (Docker Desktop) or --network=host (Linux). Replace docker with podman — all commands are identical. The :Z flag on volume mounts is required for SELinux-enabled systems (Fedora, RHEL); Docker Desktop users on macOS/Windows can omit it.
If you mount a config file, save it as /path/to/appsettings.json and mount it to /app/appsettings.json.
# Linux/macOS
cp SqlAugur/appsettings.example.json SqlAugur/publish/appsettings.json
# Edit SqlAugur/publish/appsettings.json with your server connections
# Windows (PowerShell)
Copy-Item SqlAugur\appsettings.example.json SqlAugur\publish\appsettings.json
# Edit SqlAugur\publish\appsettings.json with your server connections
Named SQL Server connections (name → connection string). Connections identify themselves to SQL Server as SqlAugur (program_name, Extended Events client_app_name) unless the connection string sets its own Application Name.
MaxRows
1000
Maximum rows returned per query, inline or in a result file
ResultFileThresholdRows
500
read_data results over this many rows (or about 40,000 characters) are written to a CSV file instead of returned inline, avoiding oversized stdio messages. 0 disables.
ResultsDirectory
sqlaugur/results under the user's local application data folder (%LOCALAPPDATA%\sqlaugur\results on Windows)
Where result files are written. Must be an absolute path.
ResultFileRetentionDays
7
Result files older than this are deleted at startup
CommandTimeoutSeconds
30
SQL command timeout for all queries and procedures
MaxConcurrentQueries
5
Maximum number of SQL queries that can execute concurrently
MaxQueriesPerMinute
60
Maximum queries allowed per minute (token bucket rate limit)
Enable progressive tool discovery — DBA tools load on demand via 3 meta-tools instead of at startup. Reduces initial context window usage. The Enable* flags still control which toolsets are allowed.
AzureKeyVaultUri
—
Azure Key Vault URI (e.g., https://myvault.vault.azure.net/). When set, secrets from the vault are added as a configuration source using DefaultAzureCredential. Key Vault secret names use -- as a section separator (e.g., a secret named SqlAugur--Servers--prod--ConnectionString maps to SqlAugur:Servers:prod:ConnectionString).
Security Note:appsettings.json is gitignored to prevent accidental credential commits. See SECURITY.md for recommended authentication methods including Windows Authentication, Azure Managed Identity, and secure credential storage options.
Tools
The server provides 31 tools organized into toolsets. Six core tools are always available. Additional toolsets are loaded at startup (static mode) or on demand (dynamic mode).
Core Tools
Tool
Description
list_servers
Lists available SQL Server instances configured in appsettings.json.
list_databases
Lists all databases on a named server with names, IDs, states, and creation dates.
read_data
Executes a read-only SQL SELECT query. Only SELECT and WITH (CTE) queries are allowed. Small results are returned as JSON; large ones go to a CSV file, returning its path, row count, columns and a 10-row preview. Pass toFile: true to always get a file.
get_query_plan
Returns the estimated or actual XML execution plan for a SELECT query.
Comprehensive table metadata in Markdown: columns, data types, nullability, defaults, identity, computed expressions, indexes, FKs, constraints.
Tool
Description
list_programmable_objects
Lists views, stored procedures, functions, and triggers. Filterable by type and schema.
get_object_definition
Returns the source definition (CREATE statement) of a programmable object.
get_extended_properties
Reads extended properties (descriptions, metadata) on tables, columns, and other objects.
get_object_dependencies
Shows what an object references and what references it — upstream and downstream dependency graphs.
Tool
Description
get_plantuml_diagram
Generates a PlantUML ER diagram with tables, columns, PKs, and FK relationships. Saves to a .puml file. Supports compact mode, schema/table filtering, and a configurable table limit (max 200).
get_mermaid_diagram
Generates a Mermaid ER diagram with tables, columns, PKs, and FK relationships. Saves to a .mmd file. Supports compact mode, schema/table filtering, and a configurable table limit (max 200).
DBA Diagnostic Tools
Each toolkit is enabled independently via config flags and requires the corresponding stored procedures installed on the target SQL Server.
All DBA tools apply response size optimisation by default — XML query plan columns are excluded and long string values are truncated to keep responses within AI context window limits. Every tool supports these optional parameters:
Parameter
Description
verbose
Return all columns with no truncation.
includeQueryPlans
Include XML execution plan columns in the output.
maxRows
Maximum rows to return per result set. Available on tools with variable-length output: BlitzIndex, BlitzLock, HealthParser, LogHunter (default 200), IndexCleanup, QueryReproBuilder.
Some tools have additional parameters: includeXmlReports (BlitzLock, HealthParser, HumanEventsBlockViewer), compact (sp_WhoIsActive), verboseMetrics (QuickieStore).
Deadlock analysis from the system_health extended event session.
sp_blitz_plan_compare
Cross-server query plan comparison — captures a plan snapshot on one server and compares it to the cached plan on a second server without using linked servers. Requires the demon_hunters branch until merged to main.
Monitors active sessions and queries — wait info, blocking details, tempdb usage, resource consumption.
Progressive Discovery
When EnableDynamicToolsets is true, only core tools load at startup. Three meta-tools let the AI discover and enable additional toolsets on demand, reducing initial context window usage:
Tool
Description
list_toolsets
Lists available toolsets with status (available, enabled, not configured) and tool counts.
get_toolset_tools
Returns detailed tool and parameter info for a specific toolset before enabling it.
enable_toolset
Enables a toolset, making its tools available. Only works if the admin has enabled the toolset via the corresponding Enable* config flag.
Example flow:
AI calls list_toolsets — sees first_responder_kit is "available" (configured but not yet enabled)
AI calls get_toolset_tools("first_responder_kit") — reviews the 7 tools and their parameters
AI calls enable_toolset("first_responder_kit") — the 7 tools are now registered and usable
AI calls sp_blitz — runs the health check as normal
In static mode (EnableDynamicToolsets: false), all enabled toolsets load at startup and the discovery tools are not registered. Schema Exploration and Diagrams toolsets are always loaded regardless of mode.
Known limitation: Progressive discovery relies on the MCP notifications/tools/list_changed notification to inform clients that new tools have been registered. Claude Code does not currently handle this notification (anthropics/claude-code#4118), so dynamically enabled toolsets will not appear. Use static mode (EnableDynamicToolsets: false) when using Claude Code.
Security
Query Validation
Every query is parsed into an Abstract Syntax Tree (AST) using Microsoft's official TSql180Parser and must pass these rules:
Single statement only — multiple statements are rejected
SELECT only — INSERT, UPDATE, DELETE, DROP, EXEC, CREATE, ALTER, and all other statement types are blocked
No SELECT INTO — prevents table creation via SELECT
No external data access — OPENROWSET (all variants including BULK, Cosmos DB, and internal), OPENQUERY, OPENDATASOURCE, OPENXML blocked
No linked servers — four-part name references are rejected
No MAXRECURSION hint — prevents overriding the default recursion limit
Cross-database queries are allowed — three-part names work by design; the security boundary is the server, not the database. To restrict to a single database, limit the login's permissions.
Because validation operates on the parsed AST, it correctly handles edge cases that defeat string-based approaches: keywords inside comments, string literals, nested block comments, and encoding tricks.
Parameter Blocking
Diagnostic stored procedures execute via whitelisted procedure names with blocked parameters that prevent writes:
First Responder Kit — all @Output* parameters blocked (prevents writing results to server tables)
DarlingData — logging and output parameters blocked (prevents table creation and data retention)
All tool executions are subject to concurrency limiting (MaxConcurrentQueries, default 5) and throughput limiting (MaxQueriesPerMinute, default 60). Excess requests are rejected with a retry message.
Connection Security
Use Windows Authentication or Azure Managed Identity where possible to avoid storing credentials in config files. When SQL Authentication is required, use environment variable overrides to inject credentials at runtime. See SECURITY.md for detailed guidance including credential stores and connection string encryption.
Known Risks
This project depends on the official Microsoft MCP C# SDK (ModelContextProtocol NuGet package, version 1.3.0). As the MCP framework handles all protocol I/O, any vulnerability in it directly affects this application's security boundary. Monitor the package for updates and upgrade when new versions are released.
The data returned from a SQL Server query could include malicious prompt injection targeting AIs. This is a risk of all AI use and cannot be mitigated by this project. Ensure you're following best practices for AI security and only connecting to trusted data sources.
Contributing
Contributions are welcome. See CONTRIBUTING.md for architecture details, development setup, testing instructions, and guidelines for adding new tools.
Noch keine Tool-Liste. Remote-Server erhalten eine nach dem ersten Health-Check.
FAQ
Fragen
Wie installiere ich SqlAugur?
Kopieren Sie auf dieser Seite die stdio-Konfiguration für Cursor, Claude Desktop, Claude Code oder VS Code. Sie nutzt die Installation des Herausgebers, keinen PlugMCP-Proxy.
Benötigt SqlAugur einen API-Schlüssel?
In diesem Eintrag ist keine erforderliche Umgebungsvariable aufgeführt.
Welchen Transport verwendet SqlAugur?
SqlAugur läuft lokal über stdio.
Wann wurde SqlAugur zuletzt geprüft?
Zuletzt geprüft: 2026-11-09. MCP server providing AI assistants with safe, read-only SQL Server access