sql-safe MCPサーバー

sql-safe-mcp

A PII-safe, read-only MCP server for coding agents working with SQL Server, MySQL and MariaDB: schema knowledge and safe queries without write-capable tools or plaintext exposure of configured PII columns.

Read the data. Protect the identity.

Give a coding agent the schema knowledge it needs to write correct code - servers, databases, tables, columns, keys, indexes, and stored procedures. On aliases you mark all_pii_safe, it can also query real rows while configured personal data stays hidden behind authenticated tokens.

PII-safe by default. Read-only by design.

• PII-safe by default - on an all_pii_safe server, the columns you configure come back as alias-bound, authenticated tokens (pii:v1:...), never as plaintext. An agent can still project, count, and filter on them with = and IN using tokens it was given, so it can follow a record without reading it. Tokens do not work on another server alias or with another key.
• Read-only by construction - seven tools, all annotated read-only. No tool writes data, and the server never executes SQL an agent wrote: metadata comes from SQLAlchemy Inspector and fixed catalog queries, and execute_sql runs only a validated, regenerated SELECT.
• Multiple servers, isolated policies - one MCP process can expose multiple named SQL Server, MySQL, and MariaDB instances. Each alias has its own connection, access level, and, when enabled, PII key and protection rules. Every database operation targets an explicit alias; there is no network discovery.

Operational safeguards

• Fails closed - SQL validation is an allowlist. Unknown syntax, unresolved lineage, and unsupported protected-value types are refused, not guessed at. The verification evidence is in the security model.
• Least access first - access_level: metadata (the default) exposes database navigation and table structure; meta_and_code additionally exposes stored procedures; execute_sql needs an explicit all_pii_safe alias with its own key. Database permissions stay the primary control, so use a least-privilege login.
• Secrets stay out of sight - connection URLs live in YAML with ${NAME} placeholders resolved from the environment. They never appear in logs or model-visible errors.
• Compact, predictable output - object-rooted results with stable sorting, literal case-insensitive name filters, and stored procedure lists that do not expand definitions.
• Errors an agent can act on - an ambiguous name lists the candidate schemas. Errors never contain connection details, credentials, keys, tokens, or rows.
• Tested against attacks, not just examples - the SQL and PII boundary is checked with an adversarial corpus of hostile statements, property-based tests (token isolation, tamper resistance), a live attack run against a really writable login with before/after snapshots, and mutation testing. Results are in Verification of the SQL boundary.
• Diagnosable failures - a timeout or connection error carries a Reference, and the stderr log records the connection stage, elapsed time, and driver error for the same reference, with credentials removed. See Logging.

Tool	Access	Purpose
list_servers	🟢 read	Configured server aliases
list_databases	🟢 read	Databases visible to the credentials
list_tables	🟢 read	Base tables, filtered by schema or name
get_table_definition	🟢 read	Columns, keys, constraints, and indexes of one table
list_stored_procedures	🟢 read	Stored procedures, without definitions; requires meta_and_code or all_pii_safe
get_stored_procedure	🟢 read	The definition of one stored procedure; requires meta_and_code or all_pii_safe
execute_sql	🟢 read	One restricted SELECT on an all_pii_safe server; protected columns return tokens

[!NOTE] Status: SQL Server supports every tool. MySQL and MariaDB (engine: mysql or mariadb, mysql+pymysql URLs) support every tool too. schema is always null there because the database is the catalog, and execute_sql uses LIMIT instead of TOP. See architecture.md.

Quick start

Install

uvx sql-safe-mcp

or

pip install sql-safe-mcp

Pin a version when you want a fixed surface: uvx sql-safe-mcp==2.0.0.

Requires Python 3.12–3.14, uv (or pip), and Microsoft ODBC Driver 18 for SQL Server when you connect to SQL Server. MySQL and MariaDB use the bundled PyMySQL driver and need nothing else.

Verified against SQL Server 2022, MySQL 8.4, and MariaDB 11.4 (see checks.md).

Install sql-safe-mcp as a stdio MCP server.
Ask whether to install it for this project/workspace or at user level. Then:
detect and preserve the host's existing MCP config;
configure the server to run via uvx sql-safe-mcp;
create a YAML config template with env placeholders only;
add SQL_SAFE_MCP_CONFIG;
report the env variables the user must set.
Do not store secrets or connection data in tracked files or output. Validate with
sql-safe-mcp --check-config once variables are available.

Configure server

Copy sql-safe-mcp.example-simple.yaml to sql-safe-mcp.yaml. This smallest configuration exposes schema metadata from one SQL Server instance and does not allow row queries. Keep the complete connection URL in an environment variable:

version: 1
servers:
  reporting:
    engine: sqlserver
    access_level: metadata
    connection_url: "${REPORTING_SQL_URL}"

Here reporting is the value an agent passes as server. A missing variable, or an engine that does not match the URL dialect, stops the MCP server at startup. For multiple servers, local PII rules, shared PII rule sets, logging, and runtime limits, use the commented examples in the configuration reference.

Check the configuration

After defining every environment variable referenced by the YAML file, point the server at it with SQL_SAFE_MCP_CONFIG (or --config) and check it without connecting to any database:

SQL_SAFE_MCP_CONFIG=sql-safe-mcp.yaml uvx sql-safe-mcp --check-config

Configuration is validated at startup, and an error names the problem without printing a URL or secret. Keep credentials in the host's own configuration and never commit them. The server acts with the database account's permissions, so use a dedicated login with the least access the job needs.

Connect an MCP client

Every client configuration needs SQL_SAFE_MCP_CONFIG plus the environment variables referenced by your YAML file. Keep connection URLs and PII keys in the MCP host's environment or configuration, never in the YAML file or other tracked files.

claude mcp add --env SQL_SAFE_MCP_CONFIG=/path/to/sql-safe-mcp.yaml --env REPORTING_SQL_URL=mssql+pyodbc://... --transport stdio sql-safe -- uvx sql-safe-mcp

Put at least one other option between the last --env and the server name, as above. Otherwise, the CLI reads the name as another KEY=value pair.

Add the server to claude_desktop_config.json:

{
  "mcpServers": {
    "sql-safe": {
      "command": "uvx",
      "args": ["sql-safe-mcp"],
      "env": {
        "SQL_SAFE_MCP_CONFIG": "/path/to/sql-safe-mcp.yaml",
        "REPORTING_SQL_URL": "mssql+pyodbc://..."
      }
    }
  }
}

codex mcp add sql-safe --env SQL_SAFE_MCP_CONFIG=/path/to/sql-safe-mcp.yaml --env REPORTING_SQL_URL=mssql+pyodbc://... -- uvx sql-safe-mcp

Use command uvx, argument sql-safe-mcp, and set SQL_SAFE_MCP_CONFIG plus every environment variable referenced by the configuration file. The server writes MCP messages to stdout and logs only to stderr.

PII-safe queries

execute_sql on the billing alias allows one restricted SELECT. For example, these arguments:

{
  "server": "billing",
  "database": "Billing",
  "sql": "SELECT CustomerId, Email FROM dbo.Customers"
}

can return a row such as [42, "pii:v1:..."]: the agent can follow the customer without reading the email address. A token is accepted only in = and IN predicates on the same server alias. If FirstName is protected, SELECT FirstName FROM Users WHERE FirstName = 'Alice' is rejected before it reaches the database. Filter protected columns only with authentic pii:v1:... tokens returned by execute_sql for the same alias; a matching prefix alone is not enough. Each token is authenticated and alias-bound, then validated and converted to a bound database value. Protection covers the columns you list, so list every column that holds personal data.

Hardening for sensitive data

An all_pii_safe alias requires verified TLS in its connection_url and fails at startup without it; require_tls: false turns the check off and logs a warning at every start (Verified TLS). Every other control below is opt-in:

• Protect by default - pii.mode: deny-by-default protects every column not listed in pii.exposed_columns, and pii.deny rejects any query that references a listed column (Deny-by-default).
• Check rules against the schema - sql-safe-mcp --verify-pii-rules [--strict] reports what each rule matches and flags unprotected columns whose names look personal (Verifying PII rules).
• Limit token reuse - rules with different domain values reject each other's tokens, and runtime.token_ttl_seconds expires old tokens (Token domains, Token lifetime).
• Keep the key out of the environment - pii_key_file reads the key from a file whose owner and permissions are checked at startup (Key file).
• Suppress small counts - runtime.min_count_cell returns small nonzero COUNT(*) values as null (Runtime limits).
• Audit reads - logging.audit: true logs the tables and columns each execute_sql call read, without SQL, values, or rows (Audit log).

Tokens hide values but are pseudonyms, not anonymization: what a token still reveals is listed in the security model. For medical deployments, see Health data of minors.

Documentation

• Configuration reference - aliases, connection URLs, PII rules, runtime limits, and logging.
• Tool reference - arguments, responses, errors, and the accepted SQL subset.
• Security model - trust boundaries, guarantees, limitations, and verification.
• Health data of minors - requirements, mechanisms, and an operator checklist for medical deployments.
• Checks - local, integration, and security verification commands.

Security

The MCP caller, SQL input, database metadata, rows, and tokens are untrusted. Database permissions remain the primary authorization control - this server never widens them. Read the full security model, including its limitations and verification evidence. To report a vulnerability, use the private channel in SECURITY.md.

Contributing

Setup, checks, the test commands, the branch and commit conventions, and the release model are in CONTRIBUTING.md. Changes that affect someone running the server are recorded in CHANGELOG.md.

License

MIT.