winget install --id=atani.mysh -e MySQL connection manager with SSH tunnel support, encrypted password storage, automatic output masking for AI and non-TTY environments, native MySQL driver support, and export formats including markdown, CSV, JSON, and PDF.
English | 日本語
Stop leaking production data to AI coding agents.
mysh is a safer MySQL CLI for the AI coding era: it manages connection profiles, SSH tunnels, and query output so Claude Code, Cursor, shell scripts, and MCP clients can work with production-like databases without accidentally seeing raw sensitive data.
Pronounced "my-sh" (/maɪʃ/), like "my shell".
> New to mysh? See the Getting Started guide for non-engineers — set up in 5 minutes and start querying with Claude Code. > > If you believe AI agents should never receive raw production PII, consider starring the repo to help more teams find it.
Traditional MySQL clients are great for humans, but risky in AI-assisted workflows: query results can be copied into prompts, captured by scripts, or returned to agents without a final privacy check.
mysh is built for that gap:
go installWhen mysh detects production data or output captured by an AI agent/script, it masks sensitive values before they leave the CLI.
mysh run production -e "SELECT id, name, email, phone FROM users LIMIT 2" --format markdown
Example masked output:
| id | name | phone | |
|---|---|---|---|
| 1 | A*** | a***@example.com | 0*** |
| 2 | B*** | b***@example.com | 0*** |
Production --raw output requires an interactive TTY confirmation, so AI tools and non-interactive scripts cannot silently bypass masking.
That means the default failure mode changes from "an agent captured raw rows" to "sensitive-looking values were masked before leaving the CLI."
See Using mysh safely with AI coding agents for recommended configuration and team workflow guidance.
mysh ships a built-in Model Context Protocol server, so AI coding agents can query your databases as a first-class tool instead of shelling out to the CLI. The same masking rules apply, and raw output cannot be requested over MCP — sensitive values are masked before they ever reach the agent.
1. Configure a connection first. The MCP server reads the same connections you set up for the CLI (~/.config/mysh/connections.yaml); it does not define its own. Set one up with mysh add, mysh import, or the Redash flags, and verify with mysh list / mysh ping. There are no add/edit/remove tools over MCP by design.
2. Register the server with Claude Code:
claude mcp add mysh -- mysh mcp
Or add it to any MCP client config (e.g. Cursor, claude_desktop_config.json):
{
"mcpServers": {
"mysh": { "command": "mysh", "args": ["mcp"] }
}
}
3. Make the master password available. Connections with encrypted credentials need the master password to decrypt them. Since the MCP transport is non-interactive, the server never prompts: run any mysh command interactively once (it is saved to the macOS Keychain / Windows Credential Manager), or set MYSH_MASTER_PASSWORD in the server's environment. If it is missing, mysh_query/mysh_ping return a clear error instead of hanging.
Exposed tools: mysh_list_connections, mysh_query, mysh_tables, mysh_ping. No hosting, API key, or extra service is required — the server runs locally over stdio. See the MCP Server Guide for details.
| Feature | mysql CLI | mycli | DBeaver | mysh |
|---|---|---|---|---|
| CLI-first workflow | ✅ | ✅ | ❌ | ✅ |
| Connection profile management | ❌ | ⚠️ | ✅ | ✅ |
| SSH tunnel management | ❌ | ❌ | ✅ | ✅ |
| Automatic masking for AI/non-TTY output | ❌ | ❌ | ❌ | ✅ |
Production --raw safety confirmation | ❌ | ❌ | ❌ | ✅ |
| Built-in MCP server for AI agents | ❌ | ❌ | ❌ | ✅ |
| Import from GUI database clients | ❌ | ❌ | — | ✅ |
| Team-safe config export without passwords | ❌ | ❌ | ⚠️ | ✅ |
| Encrypted local credential storage | ❌ | ⚠️ | ✅ | ✅ |
| Markdown/CSV/JSON/PDF export | ❌ | ❌ | ✅ | ✅ |
| MySQL 4.x old_password support | ⚠️ | ⚠️ | ⚠️ | ✅ |
mysh mcp) so AI agents can query databases natively, with masking enforcedbrew tap atani/tap
brew install mysh
Download mysh-windows-x64.msi (or -arm64) from the
latest release and run it.
The installer places mysh.exe under C:\Program Files\mysh by default (or a
folder you choose) and adds it to the system PATH.
winget install atani.mysh
winget downloads and runs the MSI for you. If you need an offline installer or want to choose the install location manually, use the MSI installer above.
Download mysh-windows-amd64.exe from the
latest release, rename it to
mysh.exe, and place it in a directory on your PATH.
> [!NOTE]
> The MSI and winget package are not code-signed yet, so Windows SmartScreen may
> show an "unknown publisher" prompt, and some endpoint-protection or ASR
> (Attack Surface Reduction) policies may block the installer. In managed
> environments an administrator may need to allow mysh.exe. A consistent
> install path (C:\Program Files\mysh) makes a path-based ASR exclusion
> straightforward, which is why the MSI keeps this path uniform across machines.
go install github.com/atani/mysh@latest
# Add a connection interactively
mysh add
# Test connection (name optional if only one connection)
mysh ping
# Connect
mysh connect
mysh [arguments]
# Add a new connection (interactive)
mysh add
# List all connections
mysh list
# Edit an existing connection
mysh edit production
# Remove a connection
mysh remove production
Connection name can be omitted when only one connection exists.
Import connections from other database tools:
mysh import --from dbeaver # Import from DBeaver
mysh import --from sequel-ace # Import from Sequel Ace
mysh import --from workbench # Import from MySQL Workbench
mysh import --from dbeaver --all # Import all without selection prompt
Export connections for team members (passwords are always excluded):
# Export all connections
mysh export > connections.yaml
# Export a specific connection
mysh export production > prod.yaml
Recipients import the shared file and enter their own password:
mysh import --from yaml --file prod.yaml
For direct DB connections where each team member has their own DB username, use --ask-user or pass one username with --db-user. If several DB connections share the same password, --reuse-password asks for it once:
mysh import --from yaml --file team-db.yaml --all --ask-user
mysh import --from yaml --file team-db.yaml --all --db-user alice --reuse-password
Exported files include environment, SSH, and mask settings, so non-engineer users get a fully configured connection — they only need to enter their own database credentials or Redash API key.
Query production databases through Redash instead of direct DB connections. No database credentials or SSH tunnels needed — just a Redash API key.
# Add a Redash connection
mysh add --name prod --redash-url https://redash.example.com --redash-key YOUR_API_KEY --redash-datasource 1
# Query through Redash (masking is applied automatically)
mysh run prod -e "SELECT * FROM users LIMIT 10"
# Test connectivity
mysh ping prod
This is ideal for non-engineer users (CS, marketing, PM) who need safe access to production data via AI assistants like Claude Code. The query flows through Redash, and mysh applies masking rules to the results before returning them.
# Test connection
mysh ping production
# Open an interactive MySQL session
mysh connect production
# Execute a SQL file
mysh run production query.sql
# Execute inline SQL
mysh run production -e "SELECT COUNT(*) FROM users"
# Show tables
mysh tables production
By default, connect and run open an ad-hoc SSH tunnel that closes when the command finishes.
For repeated access, start a persistent background tunnel:
# Start a background tunnel
mysh tunnel production
# List active tunnels
mysh tunnel
# connect/run will automatically reuse the background tunnel
mysh run production -e "SHOW PROCESSLIST"
# Stop a tunnel
mysh tunnel stop production
Multiple tunnels can run simultaneously for different connections.
mysh can automatically mask sensitive columns (email, phone, etc.) in query output. This is designed to prevent personal data from leaking into AI tool contexts.
Masking is controlled by two factors:
env in config)| env | Terminal (human) | Piped/captured (AI) |
|---|---|---|
| production | Auto-masked | Auto-masked |
| staging | Raw | Auto-masked |
| development | Raw | Raw |
connections:
- name: production
env: production
mask:
columns: ["email", "phone", "password_hash"]
patterns: ["*address*", "*secret*"]
ssh: ...
db: ...
# Force masking (even in terminal)
mysh run production --mask -e "SELECT * FROM users LIMIT 5"
# Force raw output (requires interactive confirmation for production)
mysh run production --raw -e "SELECT * FROM users LIMIT 5"
For production connections, --raw requires interactive confirmation at the terminal. Non-TTY processes (AI tools, scripts) cannot bypass masking.
| Type | Original | Masked |
|---|---|---|
| alice@example.com | a***@example.com | |
| Phone | 090-1234-5678 | 0*** |
| Name | Alice | A*** |
| Short value | ab | *** |
| NULL | NULL | NULL |
Extract specific records from a database as INSERT statements. Useful for creating reproducible test data or migrating individual records.
# Extract records matching a condition
mysh slice production products --where "category='electronics'"
# Save to file
mysh slice production products --where "id IN (7,8)" -o subset.sql
# Disable masking (requires interactive confirmation)
mysh slice production customers --where "id=3" --raw
Output example:
-- mysh slice: products WHERE category='electronics'
-- Generated at: 2026-03-15T12:00:00+09:00
INSERT INTO `products` (`id`, `name`, `price`) VALUES (7, 'Widget Pro', 2980);
INSERT INTO `products` (`id`, `name`, `price`) VALUES (8, 'Gadget Mini', NULL);
Masking rules from the connection config are always applied by default, regardless of environment. Use --raw to disable (requires interactive confirmation).
Export query results as markdown, CSV, JSON, or PDF.
# Markdown table
mysh run production -e "SELECT * FROM users LIMIT 5" --format markdown
# CSV file
mysh run production -e "SELECT * FROM users" --format csv -o users.csv
# JSON output
mysh run production -e "SELECT * FROM users LIMIT 5" --format json
# JSON file
mysh run production -e "SELECT * FROM users" --format json -o users.json
# PDF report
mysh run production -e "SELECT * FROM users" --format pdf -o report.pdf
# tables command also supports format/output
mysh tables production --format csv -o tables.csv
Supported formats: plain (default), markdown (md), csv, json, pdf
Save .sql files in ~/.config/mysh/queries/ and list them with:
mysh queries
Connections are stored in ~/.config/mysh/connections.yaml.
connections:
- name: production
env: production
ssh:
host: bastion.example.com
port: 22
user: deploy
key: ~/.ssh/id_ed25519
mask:
columns: [email, phone]
patterns: ["*address*"]
db:
host: 127.0.0.1
port: 3306
user: app
database: myapp_production
password:
- name: legacy-db
env: production
db:
host: 10.0.0.5
port: 3306
user: app
database: legacy_production
password:
driver: native # MySQL 4.x old_password support
- name: local
env: development
db:
host: localhost
port: 3306
user: root
database: myapp_dev
The driver field selects how mysh connects to MySQL.
| driver | Description | Supported versions |
|---|---|---|
cli (default) | Delegates to mysql/mycli CLI | MySQL 5.1+ |
native | Direct connection via Go's database/sql | MySQL 4.x+ |
The native driver uses go-sql-driver/mysql with allowOldPasswords=true to support MySQL 4.x old_password (mysql323) authentication. The connect command provides a simple REPL instead of mycli/mysql.
| Variable | Description |
|---|---|
MYSH_MASTER_PASSWORD | Master password for credential decryption. Useful for non-interactive contexts (AI assistants, scripts). |
The master password lookup order is: OS credential store (macOS Keychain / Windows Credential Manager) → MYSH_MASTER_PASSWORD → interactive prompt.
On macOS and Windows, the master password is saved to the OS credential store on first use, so you don't need to set MYSH_MASTER_PASSWORD for unattended runs (e.g. from AI assistants). On platforms without a supported store, use the environment variable or the interactive prompt.
Prompt instructions are useful, but they are not a data-loss prevention boundary. If raw query output is copied into an agent context, logs, CI output, or a bug report, the leak has already happened. mysh masks configured sensitive fields before the output leaves the CLI or MCP server.
Generic SQL MCP servers make database access convenient, but many are optimized for query execution rather than privacy-by-default output handling. mysh exposes MySQL access through MCP while reusing the same connection profiles, SSH tunnels, and masking rules as the CLI. Raw output cannot be requested over MCP.
No. mysh reduces accidental CLI-output leakage in AI-assisted workflows; it is not a replacement for access control, audit logging, database permissions, or a full compliance program. Treat it as a safer local workflow layer.
MYSH_MASTER_PASSWORD env var, then prompt)0600 permissions--raw on production requires interactive TTY confirmation (AI tools cannot bypass)connect limitations: Unlike mycli/mysql CLI, the built-in REPL has no tab completion, syntax highlighting, or pager. For complex interactive work, prefer run -e.allowOldPasswords: This depends on the driver's support, which may be removed in future driver updates.Japanese translations are available under docs/ja/（日本語ドキュメントは docs/ja/ を参照）.
golang.org/x/crypto - Argon2id key derivationgolang.org/x/term - Secure password input and TTY detectiongopkg.in/yaml.v3 - Configuration file parsinggithub.com/go-sql-driver/mysql - Native MySQL driver (old_password support)github.com/go-pdf/fpdf - PDF outputIf mysh helps you keep raw production data out of AI agent contexts, please ⭐ star the repo — it helps other teams find safer AI-assisted database workflows.