Mysql MCP Server

MySQL MCP Server (mysql-mcp)

📚 Full Documentation (Wiki) • Changelog • Security • Release Article

💎 Value Proposition

Integrate your AI agents with a production-ready MySQL environment. Deploy MCP v2 stateless architecture via HTTP transport. Optimize token efficiency via Code Mode. Secure your database with OAuth 2.1.

🎯 Leverage Core Benefits

---

🤖 Automate Tasks with Guided Workflows

This server includes intelligent prompts for guided workflows:

Note: This is a subset of available prompts. Use mysql_tool_index to discover the full list of guided workflows.

---

📊 Improve Observability with Resources

This server exposes a comprehensive set of resources for database observability and telemetry:

---

🚀 Deploy Your AI Database

Meet Prerequisites

• Node.js >= 24.0.0 (See .node-version for the pinned local development version)
• Bun (for executing repository automation scripts)
• MySQL server
• pnpm

Install the Server

Install via NPM / PNPM

pnpm add -g @neverinfamous/mysql-mcp

Run the server:

mysql-mcp --transport stdio --mysql "mysql://mcp_user:secure_password@localhost:3306/testdb"

Or use npx without installing:

npx @neverinfamous/mysql-mcp --transport stdio --mysql "mysql://mcp_user:secure_password@localhost:3306/testdb"

Run with Docker

Note on Namespaces: The Docker image uses the writenotenow namespace. The GitHub repo uses neverinfamous.

Linux Users: For host.docker.internal on Linux, run the container with --add-host host.docker.internal:host-gateway.

docker run -i --rm -v ./data:/app/data -v ./logs:/var/log/mysql-mcp writenotenow/mysql-mcp:latest \
  --transport stdio \
  --allowed-io-roots /app/data \
  --mysql "mysql://mcp_user:secure_password@host.docker.internal:3306/testdb"

Enable Observability via Docker Compose

Launch the full observability stack using the included template. This includes Datadog, OpenTelemetry, Prometheus, and Grafana. This spins up the MCP server, MySQL database, Redis, and observability sidecars:

cd examples/basic-mysql-datadog
cp .env.example .env
docker compose up -d

• Grafana: Available at http://localhost:3001 (Dashboard pre-loaded).
• Prometheus: Available at http://localhost:9090.
• MCP Server: Available at http://localhost:3000.

Follow Datadog & OpenTelemetry Observability Rules

Datadog Constraints:

• Authentication: Use pup CLI for authentication.
• Autodiscovery: Avoid duplicate autodiscovery configurations. Add standard tracking labels (tags.datadoghq.com/env, tags.datadoghq.com/service, tags.datadoghq.com/version) to your containers.
• WSL2 Hardening Limits: Enforce stop_grace_period: 30s, mem_limit: 1536m, and OpenMetrics timeouts of 10s.
• Host Configuration: Use DD_HOSTNAME and native /etc/docker/daemon.json cgroup configurations ("default-cgroupns-mode": "host"). Disable DD_EXTRA_PERFORMANCE_METRICS.
• Windows File Tailing: Be warned that Windows inotify tailing issues may occur with Docker/WSL2 log mapping.

OpenTelemetry Rules:

• Semantic Conventions: Enforce gen_ai.* semantics.
• Instrumentation: Use auto-instrumentation when possible.
• Context Propagation: Ensure traceparent and tracestate propagation.
• Processing & Logging: Utilize batch processors and ensure logs are formatted as JSON logs.

Audit Architecture & Exporter Healthcheck:

• The primary MCP server writes execution traces to mcp-audit.jsonl via the --audit-log flag.
• The metrics exporter runs as an independent sidecar process. It reads from mcp-audit.jsonl using the AUDIT_LOG_PATH environment variable to compute metrics.
• The exporter isolates its own internal logs by setting its --audit-log flag to exporter-audit.jsonl.
• The metrics server and exporter share a single process. Both operate on port 3000. This prevents port contention.
• Exporter Healthcheck: wget --spider -q http://127.0.0.1:3000/metrics

Build From Source

git clone https://github.com/neverinfamous/mysql-mcp.git
cd mysql-mcp
pnpm install
pnpm run build
node dist/cli.js --transport stdio --mysql "mysql://mcp_user:secure_password@localhost:3306/testdb"

---

🏗️ Use Infrastructure Templates

We provide production-ready Docker Compose templates in the examples/ directory. Use these to spin up test environments quickly. These templates include full Datadog observability out-of-the-box (Metrics, APM, and Logs).

• Basic Template: A lightweight, single-node MySQL setup. Perfect for quick starts.
• Enterprise HA Template: A highly-available InnoDB Cluster with MySQL Router and ProxySQL. Perfect for staging or production replicas.
• Full Observability Ecosystem: A complete enterprise monitoring stack with Datadog, Prometheus, Grafana, Loki, and Alloy. See the Observability Wiki for deep-dive metrics configuration.

See the README in each directory for quick start instructions.

---

⚡ Optimize Token Usage with Code Mode

Code Mode (mysql_execute_code) reduces token usage by consolidating operations in a secure sandbox.

Code executes securely in a C++ V8 isolate sandbox. It enforces strict heap limits and synchronous termination. Native wrappers map all API calls across the boundary. This guarantees defense-in-depth and fleet-standard restrictions:

Enforce Engine-Level Restrictions

• ✅ Strict V8 Isolate Boundary — Code executes within a physically separate V8 isolate. It ensures native objects and prototypes cannot cross the boundary.
• ✅ Memory & CPU Constraints — enforced at the C++ level. This includes synchronous timeouts and strict heap limits.
• ✅ API Bindings via Reference — Injects MySQL methods securely using ivm.Reference wrappers.

Validate Code Statically

• ✅ Comprehensive blocked patterns — regex rules block require(), import(), eval(), Function, process, and __proto__. They also block filesystem/network access and system commands.
• ✅ Unicode & Comment Sanitization — Strips comments and performs NFKC normalization to prevent regex evasion.
• ✅ Configurable code input limit — prevents payload-based resource exhaustion.

Protect the Runtime

• ✅ RPC Quotas — Caps RPC API calls per execution. This prevents unbounded loops.
• ✅ Execution timeout — enforces timeouts to prevent resource exhaustion. Configurable via schema timeout.
• ✅ Egress boundary enforcement — streaming JSON.stringify serialization aborts mid-flight when exceeding size caps.
• ✅ Rate limiting — Enforces per-client limits (CODEMODE_RATE_LIMIT_MAX). Uses Redis with in-memory fallbacks.
• ✅ Readonly enforcement — Returns structured errors instead of executing write methods when readonly: true.
• ✅ Audit logging — Logs every execution with UUID, client ID, metrics, and redacted code preview.
• ✅ Admin scope — Code Mode requires admin scope when OAuth is enabled.
• ✅ Full API access — Exposes all tool groups via the mysql.* namespace.

⚡ Run Only Code Mode

Run with only Code Mode enabled. A single tool provides full capability access. See Option 1: Code Mode below for the recommended IDE configuration.

This exposes just mysql_execute_code. Agents write JavaScript against the typed SDK. They compose queries and chain operations across tool groups. They return exactly the needed data in one execution. This mirrors standard serverless edge execution patterns. It ensures fixed token costs.

[!TIP] Maximize Token Savings: Instruct your AI agent to prefer Code Mode.

"When using mysql-mcp, prefer mysql_execute_code (Code Mode) for multi-step operations. This minimizes token usage."

For maximum savings, run Code Mode as your only tool using --tool-filter codemode. See the Code Mode wiki for full API documentation.

---

🌐 Enable Remote Access via Streamable & Stateless HTTP

When to use HTTP mode: Deploy mysql-mcp as a standalone server. Multiple clients can connect remotely. Use stdio mode for local development.

Use cases for HTTP mode:

• Run the server in a network-accessible Docker container
• Deploy to cloud platforms (AWS, GCP, Azure)
• Enable OAuth 2.1 authentication for enterprise security
• Share one database connection across multiple AI clients

Use the HTTP transport for remote access:

npx -y @neverinfamous/mysql-mcp \
  --transport http \
  --server-host 0.0.0.0 \
  --port 3000 \
  --allowed-io-roots /path/to/data \
  --mysql "mysql://mcp_user:secure_password@localhost:3306/testdb"

Docker:

docker run --rm -p 3000:3000 \
  -v ./data:/app/data \
  -v ./logs:/var/log/mysql-mcp \
  -e MCP_AUTH_TOKEN=my-secret-token \
  writenotenow/mysql-mcp:latest \
  --transport http --server-host 0.0.0.0 --port 3000 --allowed-io-roots /app/data --mysql "mysql://mcp_user:secure_password@host.docker.internal:3306/testdb"

The server supports Streamable HTTP and Stateless HTTP deployments. Both stateful and stateless clients can connect:

Use Streamable HTTP (Recommended)

Modern MCP protocol — single endpoint, session-based:

Rate Limit: HTTP transport is rate limited per IP (default 100/min, configurable via MCP_RATE_LIMIT_MAX). Distribute limits across deployments via Redis using graceful in-memory fallbacks.

The server manages sessions via the Mcp-Session-Id header.

Run Statelessly

Use stateless deployments where sessions are not needed:

node dist/cli.js --transport http --server-host 0.0.0.0 --port 3000 --allowed-io-roots /path/to/data --stateless --mysql "mysql://mcp_user:secure_password@..."

In stateless mode: GET /mcp returns 405. DELETE /mcp returns 204. Each POST /mcp instantiates a stateless HTTP transport.

Access Utility Endpoints

🔐 Protect Your Data with Authentication

mysql-mcp supports two authentication mechanisms for HTTP transport:

Authenticate with Bearer Token (--auth-token)

Use lightweight authentication for development:

node dist/cli.js --transport http --server-host 0.0.0.0 --port 3000 --allowed-io-roots /path/to/data --auth-token my-secret --mysql "mysql://mcp_user:secure_password@..."

# Or via environment variable
export MCP_AUTH_TOKEN=my-secret
node dist/cli.js --transport http --server-host 0.0.0.0 --port 3000 --allowed-io-roots /path/to/data --mysql "mysql://mcp_user:secure_password@..."

Clients must include Authorization: Bearer my-secret on all requests. /health and / are exempt. Unauthenticated requests receive 401. Responses include WWW-Authenticate: Bearer headers per RFC 6750.

Authenticate with OAuth 2.1

Use full OAuth 2.1 for production deployments:

node dist/cli.js \
  --transport http \
  --server-host 0.0.0.0 \
  --port 3000 \
  --allowed-io-roots /path/to/data \
  --mysql "mysql://mcp_user:secure_password@localhost:3306/testdb" \
  --oauth-enabled \
  --oauth-issuer https://localhost:8080/realms/mysql-mcp \
  --oauth-audience mysql-mcp-client

Additional flags: --oauth-jwks-uri <url> (auto-discovered if omitted), --oauth-clock-tolerance <seconds> (default: 60).

Enforce OAuth Scopes

Enforce access control using OAuth scopes:

Ensure RFC Compliance

This implementation follows full OAuth 2.1 for production multi-tenant deployments:

• ✅ RFC 9728 Protected Resource Metadata (/.well-known/oauth-protected-resource)
• ✅ RFC 8414 Authorization Server Discovery with caching
• ✅ RFC 7591 OAuth 2.1 Dynamic Client Registration
• ✅ JWT validation with JWKS support (TTL: 1 hour)
• ✅ MySQL-specific scopes: read, write, admin, full, db:{name}, table:{db}:{table}
• ✅ Per-tool scope enforcement via AsyncLocalStorage context threading

Note for Keycloak users: Add an Audience mapper to your client. This includes the correct aud claim. (Client → Client scopes → dedicated scope → Add mapper → Audience)

[!NOTE] Per-tool scope enforcement: The server enforces scopes at the tool level. Each tool group requires a specific scope. When OAuth is enabled, every tool invocation checks the calling token's scopes before execution. The server skips scope checks entirely when OAuth is not configured.

[!WARNING] HTTP without authentication: Exposing --transport http without authentication grants unrestricted access. Always enable authentication for production HTTP deployments. See SECURITY.md for details.

⚡ Simplify AI Integration with Client Configs

Install Agent Skills

For optimal AI performance, mysql-mcp includes specialized agent instructions in the skills/ directory. These skills teach AI coding assistants (like Cline, Antigravity, or Copilot) how to effectively use the server's capabilities, tool filters, and Code Mode.

To install: Copy the .md files from the skills/ directory into your AI assistant's designated skills or custom instructions folder (e.g., .cursor/rules/, .agents/skills/, etc.). This enables your agent to autonomously navigate the mysql-mcp ecosystem without hallucinating usage patterns.

Configure IDE Settings

Add a configuration to your IDE's MCP settings file:

Option 1: Code Mode (Maximum Token Savings, 🌟 Recommended)

Best for: General MySQL AI agent tasks. Exposes mysql_execute_code for full sandboxed toolset access.

{
  "mcpServers": {
    "mysql-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "@neverinfamous/mysql-mcp",
        "--transport",
        "stdio",
        "--tool-filter",
        "codemode"
      ],
      "env": {
        "MYSQL_HOST": "localhost",
        "MYSQL_PORT": "3306",
        "MYSQL_USER": "mcp_user",
        "MYSQL_PASSWORD": "secure_password",
        "MYSQL_DATABASE": "testdb"
      },
      "timeout": 600
    }
  }
}

Option 2: Cluster (Tools for InnoDB Cluster Monitoring)

Best for: Monitoring InnoDB Cluster, Group Replication status, and cluster topology.

⚠️ Prerequisites:

• InnoDB Cluster must be configured and running with Group Replication enabled
• Connect to a cluster node directly (e.g., localhost:3307) — NOT a standalone MySQL instance
• Use cluster_admin or root user with appropriate privileges
• See MySQL Ecosystem Setup Guide for cluster setup instructions

{
  "mcpServers": {
    "mysql-mcp-cluster": {
      "command": "npx",
      "args": [
        "-y",
        "@neverinfamous/mysql-mcp",
        "--transport",
        "stdio",
        "--tool-filter",
        "cluster"
      ],
      "env": {
        "MYSQL_HOST": "localhost",
        "MYSQL_PORT": "3307",
        "MYSQL_USER": "cluster_admin",
        "MYSQL_PASSWORD": "cluster_password",
        "MYSQL_DATABASE": "mysql"
      },
      "timeout": 600
    }
  }
}

Option 3: Ecosystem (Tools for InnoDB Cluster Deployments)

Best for: MySQL Router, ProxySQL, MySQL Shell, and InnoDB Cluster deployments.

⚠️ Prerequisites:

• InnoDB Cluster requires a running cluster. This enables Router REST API authentication.
• Router REST API uses self-signed HTTPS certificates. Set MYSQL_ROUTER_INSECURE=true to bypass verification.
• X Protocol: InnoDB Cluster includes the MySQL X Plugin by default. Set MYSQL_XPORT to the Router's X Protocol port (e.g., 6448). This enables mysqlsh_import_json and docstore tools
• See MySQL Ecosystem Setup Guide for detailed instructions

{
  "mcpServers": {
    "mysql-mcp-ecosystem": {
      "command": "npx",
      "args": [
        "-y",
        "@neverinfamous/mysql-mcp",
        "--transport",
        "stdio",
        "--tool-filter",
        "ecosystem"
      ],
      "env": {
        "MYSQL_HOST": "localhost",
        "MYSQL_PORT": "3307",
        "MYSQL_XPORT": "6448",
        "MYSQL_USER": "cluster_admin",
        "MYSQL_PASSWORD": "cluster_password",
        "MYSQL_DATABASE": "testdb",
        "MYSQL_ROUTER_URL": "https://localhost:8443",
        "MYSQL_ROUTER_USER": "rest_api",
        "MYSQL_ROUTER_PASSWORD": "router_password",
        "MYSQL_ROUTER_INSECURE": "true",
        "PROXYSQL_HOST": "localhost",
        "PROXYSQL_PORT": "6032",
        "PROXYSQL_USER": "radmin",
        "PROXYSQL_PASSWORD": "radmin",
        "MYSQLSH_PATH": "mysqlsh"
      },
      "timeout": 600
    }
  }
}

Note: Port 6448 is the standard MySQL Router X Protocol port.

Customization Notes:

• Replace /path/to/mysql-mcp/ with your actual installation path. This is only needed when building from source. It is also required when running local files directly.
• Update credentials with your actual values
• For Windows: Use forward slashes (e.g., C:/mysql-mcp/dist/cli.js) or escape backslashes
• For Windows MySQL Shell: "MYSQLSH_PATH": "C:\\Program Files\\MySQL\\MySQL Shell\\bin\\mysqlsh.exe"
• Router Authentication: Router REST API authenticates against the InnoDB Cluster metadata. The cluster must be running for authentication to work.
• Cluster Resource: Connect to an InnoDB Cluster node. This unlocks the mysql://cluster resource

📖 See the Configuration Wiki for more configuration options.

---

🔗 Integrate Any MySQL Environment

Connect to MySQL on Host Machine

For native local installations, use localhost. If running the server in Docker, use host.docker.internal:

[
  "--mysql",
  "mysql://mcp_user:secure_password@host.docker.internal:3306/testdb"
]

Connect to MySQL in Another Docker Container

For Docker setups, use host.docker.internal. For Docker Compose, use the mysql service name.

Create a network and run MySQL:

docker network create mynet
docker run -d --name mysql-db --network mynet -e MYSQL_ROOT_PASSWORD=pass mysql:latest

Run MCP server on the same network:

docker run -i --rm --network mynet writenotenow/mysql-mcp:latest \
  --transport stdio --mysql "mysql://mcp_user:secure_password@mysql-db:3306/testdb"

Connect to Remote/Cloud MySQL (RDS, Cloud SQL, etc.)

Use the remote hostname directly:

[
  "--mysql",
  "mysql://mcp_user:secure_password@your-instance.region.rds.amazonaws.com:3306/testdb"
]

Tip: Ensure remote MySQL servers allow connections from Docker's IP range. Verify firewalls/security groups permit port 3306.

---

🛠️ Optimize Limits with Tool Filtering

[!IMPORTANT] AI IDEs like Cursor have tool limits. Filter tools to ensure compatibility with IDE limits. We offer many specialized tools. All shortcuts and tool groups include Code Mode by default. To exclude it, add -codemode to your filter: --tool-filter core,json,-codemode

Discover Filtering Options

The --tool-filter argument accepts shortcuts, groups, or tool names — mix and match freely:

Use Predefined Shortcuts

Filter by Tool Groups

Note: Code Mode (mysql_execute_code) is automatically added to all groups.

---

Review Syntax Reference

Legacy Syntax (still supported): If you start with a negative filter (e.g., -ecosystem), it enables all tools first. It then subtracts the specified tools.

Customize Tool Selection

List tool names without + prefix to create a custom whitelist.

The easiest way to filter is using whitelist mode. Simply specify the shortcut you want. Everything else is automatically disabled.

# Enable specific tools (whitelist mode)
--tool-filter "mysql_read_query,mysql_write_query,mysql_list_tables"

# Mix tools from different groups
--tool-filter "mysql_read_query,mysql_explain,mysql_json_extract"

# Combine with a shortcut or group
--tool-filter "starter,+mysql_spatial_distance,+mysql_json_diff"

Use this for automated clients requiring a minimal capability set.

📖 See the Tool Filtering Wiki for advanced examples.

---

🔧 Customize with Advanced Configuration

See the configuration templates in the project root for setup details.

For specialized setups, see these Wiki pages:

---

⚡ Maximize Server Performance

The server caches schema metadata to reduce repeated queries during tool/resource invocations.

Review Performance Benchmarks

The server handles millions of ops/sec across core execution paths. This ensures minimal latency and maximum throughput. Every component is tuned for enterprise-scale workloads. Enjoy sub-millisecond sandbox cold starts and optimized reverse lookups.

Benchmark Baselines:

• parseToolFilter: ~32,000-62,000 ops/sec
• CodeModeSandbox.create cold start: ~2.78M ops/sec
• Sandbox dispose: ~2.37M ops/sec
• SandboxPool init: ~109k ops/sec
• Set.has tool check: ~4.4M ops/sec
• Map.get reverse lookup: ~4.5M ops/sec
• Map.get URI match: ~5.1M ops/sec
• validateCode safe short: ~173k ops/sec
• validateCode blocked: ~298k ops/sec
• checkRateLimit: ~205k ops/sec
• sanitizeResult small payload: ~1.49M ops/sec
• prompt schema parse: ~1.3M ops/sec

Tip: Lower METADATA_CACHE_TTL_MS for development (e.g., 5000). Increase it for production with stable schemas (e.g., 300000 = 5 min).

Payload optimization: Tools support summary: true and limit parameters. These condense large payloads.

---

🎛️ Master Server Configuration

Configure CLI Options

Note: Items marked with — in the CLI flag column are configured exclusively via environment variables.

Documentation truncated — see the full README on GitHub.