Codex CLI for MySQL Development: MCP Servers, Schema Exploration, and Query Workflows on MySQL 9.7

Why MySQL Needs Its Own Article

Codex CLI articles already cover PostgreSQL, MongoDB, Redis, DuckDB, SQLite, Neo4j, and the search-engine triad. MySQL — still the world’s most deployed open-source relational database — has been conspicuously absent. With MySQL 9.7.0 LTS shipping in May 2026 with JSON Duality Views in Community Edition, the Hypergraph Optimiser, and in-database JavaScript ¹, and at least five mature MCP servers now available, the integration surface is rich enough to warrant dedicated coverage.

The MySQL MCP Server Landscape

Four servers stand out for Codex CLI integration, each occupying a different niche.

askdba/mysql-mcp-server (Go, read-only)

A single Go binary that exposes ten tools — list_databases, list_tables, describe_table, run_query, ping, server_info, list_connections, use_connection, vector_search, and vector_info — across MySQL 8.0–9.7 and MariaDB 10.x–11.x ². Queries are hard-limited to SELECT, SHOW, DESCRIBE, and EXPLAIN, making it the safest option for production introspection. Multi-DSN support lets a single server instance connect to several MySQL instances with tool-based switching.

Install via Homebrew (brew install askdba/tap/mysql-mcp-server) or Docker (ghcr.io/askdba/mysql-mcp-server:latest).

benborla/mcp-server-mysql (Node.js, configurable permissions)

A Node.js server with SSH tunnel support, connection pooling, query caching, and SQL injection prevention ³. Permissions are granular: ALLOW_INSERT_OPERATION, ALLOW_UPDATE_OPERATION, and ALLOW_DELETE_OPERATION flags control write access independently. The server annotates its mysql_query tool with MCP ToolAnnotations, hinting destructive operations to the agent so Codex CLI’s approval mode can gate them correctly.

designcomputer/mysql_mcp_server (Python, resource-based)

A Python implementation that exposes MySQL tables as MCP resources rather than just tools, letting Codex CLI browse schemas through the resource discovery protocol before issuing queries ⁴. Comprehensive logging provides an audit trail for every database operation.

awslabs/mysql-mcp-server (Python, Aurora-native)

The AWS Labs server supports two connection methods — the RDS Data API for serverless Aurora MySQL connections and direct asyncmy connections for self-hosted MySQL, RDS MySQL, RDS MariaDB, or Aurora MySQL ⁵. It integrates with AWS Secrets Manager for credential retrieval and supports read-only mode by default with configurable write access.

Google MCP Toolbox for Databases

Google’s MCP Toolbox (v1.0.0, April 2026) provides a production-grade, database-agnostic MCP server with native MySQL support, connection pooling, IAM authentication, and OpenTelemetry observability out of the box ⁶. Its prebuilt tools — list_tables, execute_sql, and list-table-stats-tool for MySQL — plus Cloud SQL integration make it the enterprise choice for GCP-hosted MySQL instances.

Configuring Codex CLI

Single-server setup (askdba, read-only)

[mcp_servers.mysql]
command = "mysql-mcp-server"
args = ["--dsn", "user:password@tcp(127.0.0.1:3306)/mydb"]

Node.js server with SSH tunnel

[mcp_servers.mysql]
command = "npx"
args = ["-y", "@benborla/mcp-server-mysql"]

[mcp_servers.mysql.env]
MYSQL_HOST = "127.0.0.1"
MYSQL_PORT = "3306"
MYSQL_USER = "codex_agent"
MYSQL_PASSWORD = ""
MYSQL_DATABASE = "mydb"
ALLOW_INSERT_OPERATION = "false"
ALLOW_UPDATE_OPERATION = "false"
ALLOW_DELETE_OPERATION = "false"

Use env_vars to forward credentials from your shell environment rather than hardcoding them:

[mcp_servers.mysql]
command = "npx"
args = ["-y", "@benborla/mcp-server-mysql"]
env_vars = ["MYSQL_HOST", "MYSQL_PORT", "MYSQL_USER", "MYSQL_PASSWORD", "MYSQL_DATABASE"]

AWS Aurora MySQL via RDS Data API

[mcp_servers.aurora_mysql]
command = "uvx"
args = ["awslabs.mysql-mcp-server"]

[mcp_servers.aurora_mysql.env]
CONNECTION_TYPE = "rds_data_api"
RESOURCE_ARN = "arn:aws:rds:eu-west-1:123456789:cluster:my-cluster"
SECRET_ARN = "arn:aws:secretsmanager:eu-west-1:123456789:secret:my-secret"
DATABASE = "mydb"
AWS_PROFILE = "dev"
READONLY = "true"

Google MCP Toolbox (Streamable HTTP)

[mcp_servers.toolbox_mysql]
url = "http://127.0.0.1:5000/mcp"

The Toolbox runs as a sidecar; configure it with a tools.yaml pointing at your Cloud SQL or self-hosted MySQL instance ⁶.

AGENTS.md for MySQL 9.7 Projects

Drop this into your repository root so Codex CLI adopts project-wide conventions:

# AGENTS.md — MySQL 9.7 Project Conventions

## Database
- MySQL 9.7 LTS (Community Edition) with InnoDB default
- Use JSON Duality Views for document-relational hybrid access patterns
- Prefer the Hypergraph Optimiser path for complex joins (9.7+ default)

## Anti-hallucination rules
- Do NOT use syntax from MySQL 5.7 or 8.0-only features that have been superseded
- `mysql_native_password` is deprecated; use `caching_sha2_password`
- `utf8` is an alias for `utf8mb3`; always specify `utf8mb4` explicitly
- Window functions require MySQL 8.0+; CTEs require MySQL 8.0+
- JSON Duality Views are 9.7+ only — do not assume availability on 8.4 LTS

## Naming conventions
- Tables: snake_case, plural (e.g. `order_items`)
- Columns: snake_case (e.g. `created_at`)
- Indexes: `idx_<table>_<columns>` (e.g. `idx_orders_customer_id`)
- Foreign keys: `fk_<table>_<ref_table>` (e.g. `fk_order_items_orders`)

## Safety
- Always include `WHERE` clauses in `UPDATE` and `DELETE` statements
- Use transactions for multi-statement writes
- Prefer `ALTER TABLE ... ALGORITHM=INSTANT` where supported
- Never run `DROP TABLE` or `TRUNCATE` without explicit confirmation

Workflow Patterns

1. Schema exploration and documentation

codex "Connect to the MySQL database via the mysql MCP server.
List all tables, describe each one, and produce a markdown
data dictionary with column types, indexes, and foreign key
relationships."

The agent calls list_tables → describe_table for each table → writes a structured markdown file. With the askdba server, server_info also captures the MySQL version and global variables for context.

2. Query generation with safety gating

codex --approval-mode suggest "Write a query that returns the
top 10 customers by lifetime order value, joining orders,
order_items, and products. Use CTEs for clarity. Run it via
the mysql MCP server and explain the EXPLAIN output."

In suggest mode, Codex CLI shows the generated SQL and waits for approval before executing via run_query. The agent can then call run_query with EXPLAIN ANALYZE prefixed to the query and interpret the execution plan.

3. Migration script generation

codex "Examine the current schema via the mysql MCP server.
Generate an Alembic migration that adds a 'status' enum column
to the orders table with values ('pending', 'processing',
'shipped', 'delivered', 'cancelled'), a default of 'pending',
and an index on (status, created_at). Use ALGORITHM=INSTANT
if the MySQL version supports it."

The agent introspects the live schema to avoid column name collisions, checks server_info for the MySQL version, and generates a migration that uses INSTANT DDL on MySQL 8.0.12+ ⁷.

4. Batch audit with `codex exec`

codex exec \
  --prompt "Connect to MySQL via the mysql MCP server. \
    List all tables missing a primary key, tables using \
    utf8mb3 instead of utf8mb4, and tables with no indexes \
    beyond the primary key. Return JSON." \
  --output-schema '{"tables_no_pk":["string"],"utf8mb3_tables":["string"],"unindexed_tables":["string"]}'

This pattern works well in CI pipelines for drift detection — the structured JSON output can feed into alerting or reporting tools.

flowchart TD
    A[codex exec] --> B[Connect to MySQL MCP]
    B --> C[list_tables]
    C --> D[describe_table × N]
    D --> E{Audit checks}
    E --> F[Missing PKs]
    E --> G[utf8mb3 charset]
    E --> H[No secondary indexes]
    F & G & H --> I[Structured JSON output]
    I --> J[CI pipeline / alerting]

Composing Multiple MCP Servers

For a typical MySQL-backed web application, compose the MySQL MCP server with complementary servers:

# Schema exploration and queries
[mcp_servers.mysql]
command = "mysql-mcp-server"
args = ["--dsn", "codex:@tcp(127.0.0.1:3306)/myapp"]

# GitHub issues and PRs
[mcp_servers.github]
url = "https://api.githubcopilot.com/mcp/"
bearer_token_env_var = "GITHUB_TOKEN"

# Filesystem access for migration files
[mcp_servers.filesystem]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-filesystem", "./migrations"]

graph LR
    subgraph "Codex CLI"
        Agent
    end
    subgraph "MCP Servers"
        MySQL["mysql-mcp-server<br/>Schema + Queries"]
        GH["GitHub MCP<br/>Issues + PRs"]
        FS["Filesystem MCP<br/>Migration files"]
    end
    Agent --> MySQL
    Agent --> GH
    Agent --> FS

This three-server stack lets the agent read a GitHub issue describing a schema change, introspect the current database state, generate a migration file, and open a pull request — all in a single turn.

Model Selection

Task	Recommended model	Rationale
Complex query generation	`gpt-5.5` (default)	Best at multi-join SQL, CTEs, window functions ⁸
Schema documentation	`gpt-5.4-mini`	Lower cost; schema introspection is structured ⁸
Batch audits via `codex exec`	`codex-mini-latest`	Optimised for fast, focused tool-calling loops ⁸
Migration review	`gpt-5.5`	Needs version-awareness for DDL algorithm selection

For ChatGPT-authenticated sessions, omit the --model flag to track the recommended default automatically.

Sandbox and Security Considerations

Network access. All MySQL MCP servers need TCP access to the database host. When using Codex CLI’s default sandbox, ensure port 3306 (or your custom port) is allowed. The full-auto approval mode will execute queries without confirmation — use suggest or auto-edit for production databases.

Credential hygiene. Never hardcode passwords in config.toml. Use env_vars to forward credentials from your shell, or the AWS Labs server’s Secrets Manager integration for cloud deployments. Create a dedicated MySQL user with minimal privileges:

CREATE USER 'codex_agent'@'%' IDENTIFIED BY '...';
GRANT SELECT, SHOW DATABASES, SHOW VIEW, PROCESS ON *.* TO 'codex_agent'@'%';
-- Add write grants only if needed:
-- GRANT INSERT, UPDATE, DELETE ON mydb.* TO 'codex_agent'@'%';

Read-only by default. Start with a read-only MCP server (askdba) or read-only flags (benborla with all ALLOW_* set to false, awslabs with READONLY=true). Escalate to write access only when the workflow explicitly requires it.

Approval gating. For write-capable servers, use Codex CLI’s approval mode:

[mcp_servers.mysql]
command = "npx"
args = ["-y", "@benborla/mcp-server-mysql"]
approval_mode = "always"

This forces a confirmation prompt before every tool invocation, including destructive queries.

Limitations

Training data lag. GPT-5.5’s training data predates MySQL 9.7 LTS (May 2026). The agent may not know about JSON Duality Views in Community Edition, the Hypergraph Optimiser, or in-database JavaScript without MCP-provided context ¹. Use an AGENTS.md file to bridge the gap.
No real-time monitoring. None of the community MCP servers expose SHOW PROCESSLIST, slow query log tailing, or Performance Schema dashboards. For operational monitoring, use Datadog or Grafana integrations separately.
Tool budget. The askdba server exposes 10 tools; composing it with other MCP servers can consume significant context. Use enabled_tools to expose only the tools needed for the current workflow.
MariaDB divergence. The askdba server supports MariaDB, but MariaDB 11.x and MySQL 9.7 have diverged significantly on JSON handling, CTEs, and window functions. Specify the exact engine version in AGENTS.md to avoid cross-engine hallucinations.
Vector search maturity. MySQL 9.0+ vector search support is still evolving; the vector_search and vector_info tools in askdba may produce unexpected results on older versions ².
⚠️ Google MCP Toolbox prebuilt MySQL tools may lag behind the latest Toolbox release. Check googleapis.github.io/genai-toolbox/reference/prebuilt-tools/ for current tool availability.

Citations

Oracle, “MySQL 9.7.0 LTS Is Now Available: Expanded Community Capabilities and Dynamic Data Masking for Enterprise,” Oracle MySQL Blog, May 2026. https://blogs.oracle.com/mysql/mysql-9-7-0-lts-is-now-available-expanded-community-capabilities-and-dynamic-data-masking-for-enterprise ↩ ↩²
askdba, “mysql-mcp-server: MySQL Server implementation for Model Context Protocol (MCP) written in Go,” GitHub, 2026. https://github.com/askdba/mysql-mcp-server ↩ ↩²
benborla, “mcp-server-mysql: A Model Context Protocol server that provides access to MySQL databases,” GitHub, 2026. https://github.com/benborla/mcp-server-mysql ↩
designcomputer, “mysql_mcp_server: A Model Context Protocol (MCP) server that enables secure interaction with MySQL databases,” GitHub, 2026. https://github.com/designcomputer/mysql_mcp_server ↩
AWS Labs, “Amazon Aurora MySQL MCP Server,” Open Source MCP Servers for AWS, 2026. https://awslabs.github.io/mcp/servers/mysql-mcp-server ↩
Google, “MCP Toolbox for Databases,” googleapis.github.io, 2026. https://googleapis.github.io/genai-toolbox/ ↩ ↩²
Oracle, “MySQL 9.7 Reference Manual: ALTER TABLE and Generated Columns,” dev.mysql.com, 2026. https://dev.mysql.com/doc/refman/9.7/en/ ↩
OpenAI, “Models — Codex,” OpenAI Developers, 2026. https://developers.openai.com/codex/models ↩ ↩² ↩³

Codex CLI for MySQL Development: MCP Servers, Schema Exploration, and Query Workflows on MySQL 9.7

Why MySQL Needs Its Own Article

The MySQL MCP Server Landscape

askdba/mysql-mcp-server (Go, read-only)

benborla/mcp-server-mysql (Node.js, configurable permissions)

designcomputer/mysql_mcp_server (Python, resource-based)

awslabs/mysql-mcp-server (Python, Aurora-native)

Google MCP Toolbox for Databases

Configuring Codex CLI

Single-server setup (askdba, read-only)

Node.js server with SSH tunnel

AWS Aurora MySQL via RDS Data API

Google MCP Toolbox (Streamable HTTP)

AGENTS.md for MySQL 9.7 Projects

Workflow Patterns

1. Schema exploration and documentation

2. Query generation with safety gating

3. Migration script generation

4. Batch audit with codex exec

Composing Multiple MCP Servers

Model Selection

Sandbox and Security Considerations

Limitations

Citations

4. Batch audit with `codex exec`