Repository avatar
Databases
v0.1.2
active

sakila-mcp

io.github.pantoniades/sakila-mcp

Read-only MySQL MCP server with a fail-closed SQL classifier and structured results.

Documentation

sakila-mcp

sakila-mcp is an MCP server that lets an LLM agent inspect and query any MySQL database, read-only, with a fail-closed SQL classifier between the agent and the server.

sakila-mcp is a general-purpose MySQL MCP server for any MySQL database. It is named after Sakila, the MySQL dolphin mascot, and is not specific to the Sakila sample database.

Independent open-source project. Not affiliated with, sponsored by or endorsed by Oracle. MySQL is a trademark of Oracle and/or its affiliates.

Status: v0.1 (0.1.1 is the first release on PyPI). What's described here works today. Expect the interface to change.

Quickstart

You need uv and a MySQL account that can only read (the grants are under Safety model).

cat > ~/.config/sakila-mcp.toml <<'EOF'
host = "db.example.internal"
user = "sakila_mcp"
password = "..."
database_allowlist = ["app_db"]
EOF
chmod 600 ~/.config/sakila-mcp.toml

claude mcp add sakila -- uvx sakila-mcp --config ~/.config/sakila-mcp.toml

Then ask "what's in app_db and how do orders relate to customers?" or "why is this query slow?". Any MCP client that speaks stdio works; see Use it.

What makes it different

  • It fails closed. A statement the classifier can't parse is refused, and so is any function that isn't a known MySQL built-in.
  • It doesn't trust one lexer. Comments are re-scanned with MySQL's own rules, because a parser and the server can disagree about where a comment ends.
  • A refusal is a result, not an error: a code, a reason and a suggestion, so the agent can rephrase instead of giving up.
  • Everything is bounded without touching your SQL: rows, cell size and execution time are capped, and no LIMIT is ever added to a query.
  • The classifier isn't the only layer. Sessions are read-only transactions, re-initialised on every checkout, and the connection can't send more than one statement.
  • It knows MySQL's release lines (LTS, Innovation, calendar versions) and reads both EXPLAIN FORMAT=JSON shapes.
  • It says where it stops: the database allowlist is a name filter, and the real boundary is the account's grants.

What it does

Six tools, all read-only:

ToolReturns
server_infoMySQL version and release line (LTS / Innovation), uptime, session time zone, key globals (buffer pool, binlog format, GTID mode, read_only, sql_mode, charset), and the sakila-mcp version.
list_databasesDatabases you're allowed to query, system schemas flagged.
list_tablesTables and views in one database: engine, approximate rows, data and index size.
describe_tableColumns, indexes (key parts in order, uniqueness, cardinality), foreign keys, engine, row format, collation.
run_queryColumns and rows for one statement, capped at max_rows. Parameters via ? placeholders.
explain_queryEXPLAIN FORMAT=JSON plus a summary (full table scans, filesort, temporary tables, largest row estimate), or FORMAT=TREE text. Doesn't run the query.

Plus one resource, mysql://capabilities: version, release line, EXPLAIN JSON format version, and whether sys is visible.

Results are structured, not prose. DECIMAL comes back as a string so it stays exact, dates and times as ISO 8601, JSON parsed, binary as <binary N bytes>. Long cells are cut at max_cell_bytes with a marker saying how big they were.

When the server says no, the agent gets a result with refused: true, a reason and a suggestion, not an error, so it can rephrase. That includes SQL the classifier can't parse. Mistakes MySQL reports (unknown table or column, wrong parameter count) and infrastructure failures are tool errors.

Safety model

Every session, on every pool checkout:

  • SET SESSION TRANSACTION READ ONLY
  • max_execution_time = statement_timeout_ms, lock_wait_timeout = 5 s
  • connection attributes program_name=sakila-mcp and sakila_mcp_version, and every statement is tagged /* sakila-mcp tool=... */, so you can find it in the processlist and performance_schema

Every statement from run_query and explain_query goes through a SQL classifier (sqlglot plus token-level checks) before it reaches MySQL. It fails closed: if it can't parse it, it refuses it.

  • Allowed: one SELECT, WITH, TABLE, VALUES, DESCRIBE, EXPLAIN, or a SHOW from a fixed list.
  • Refused: everything else, more than one statement, /*! ... */ version comments, INTO (OUTFILE, DUMPFILE, variables), locking reads (FOR UPDATE, FOR SHARE), := assignment, the SET_VAR hint, EXPLAIN ANALYZE (it executes), functions that block, lock, wait or read files (SLEEP, BENCHMARK, GET_LOCK, LOAD_FILE, ...), and any function that isn't a known built-in (stored functions can write).
  • Parameters are bound by the server (PREPARE / EXECUTE ... USING), never spliced into the SQL.
  • Results stop at max_rows; the query is cut off, not rewritten with a LIMIT.

The database allowlist is a filter on names, not an isolation boundary. information_schema, performance_schema and sys are always readable, and SHOW PROCESSLIST, SHOW ENGINE INNODB STATUS and performance_schema statement history show other sessions' SQL, literal values included. The mysql schema is denied unless you list it.

The real boundary is the account's grants. Give sakila-mcp its own account with only what it needs:

CREATE USER 'sakila_mcp'@'%' IDENTIFIED BY '...';
GRANT SELECT, SHOW VIEW ON app_db.* TO 'sakila_mcp'@'%';
-- Optional: SHOW PROCESSLIST / SHOW REPLICA STATUS. PROCESS exposes other sessions' SQL.
GRANT PROCESS, REPLICATION CLIENT ON *.* TO 'sakila_mcp'@'%';

The password never appears in tool results, logs or error text. There's a test for that.

Requirements

  • Python 3.12 or later and uv
  • MySQL 8.4 or later. CI runs the integration tests against 8.4, 9.7 and 26.10. 8.0 is tested best-effort: failures there don't block a release, and functions that exist only in 8.0 are refused as unknown.

Install

Nothing to install: uvx sakila-mcp downloads and runs the latest release. Pin a version with uvx sakila-mcp@0.1.2, or install it with uv tool install sakila-mcp.

uvx sakila-mcp --help

To run from a checkout instead, see Development.

Configure

Environment variables (SAKILA_MCP_*), a TOML file passed with --config, or both. Environment variables win over the file. Keep the password in the file, chmod 600 it, and keep it out of version control.

# ~/.config/sakila-mcp.toml
host = "db.example.internal"
port = 3306
user = "sakila_mcp"
password = "..."
database_allowlist = ["app_db"]   # empty = every non-system database
KeyEnv varDefault
hostSAKILA_MCP_HOST127.0.0.1
portSAKILA_MCP_PORT3306
userSAKILA_MCP_USER(required)
passwordSAKILA_MCP_PASSWORDempty
database_allowlistSAKILA_MCP_DATABASE_ALLOWLISTempty (all)comma-separated in the env var
max_rowsSAKILA_MCP_MAX_ROWS200rows returned per query
max_cell_bytesSAKILA_MCP_MAX_CELL_BYTES4096longer cells are truncated
statement_timeout_msSAKILA_MCP_STATEMENT_TIMEOUT_MS10000session max_execution_time
pool_sizeSAKILA_MCP_POOL_SIZE4
read_onlySAKILA_MCP_READ_ONLYtruefalse refuses to start; there is no write mode

The server connects to MySQL on the first tool call, not at startup. It starts and lists its tools even when MySQL is down; a wrong host or password shows up as a tool error on the first call (and in stderr), and the next call tries again.

Use it

Claude Code:

claude mcp add sakila -- uvx sakila-mcp --config ~/.config/sakila-mcp.toml

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "sakila": {
      "command": "uvx",
      "args": ["sakila-mcp", "--config", "/home/you/.config/sakila-mcp.toml"]
    }
  }
}

Ollama, with local models. Ollama runs the model but isn't an MCP client itself, so use ollmcp (mcp-client-for-ollama), a terminal client that connects Ollama models to MCP servers:

uv tool install ollmcp

It reads the same mcpServers format as Claude Desktop. Save this as, say, ~/.config/sakila-mcp.ollmcp.json:

{
  "mcpServers": {
    "sakila": {
      "command": "uvx",
      "args": ["sakila-mcp", "--config", "/home/you/.config/sakila-mcp.toml"]
    }
  }
}

and start it with a model that supports tool calling. ollama show <model> lists tools under Capabilities, and https://ollama.com/search?c=tools lists the models that have it:

ollmcp -j ~/.config/sakila-mcp.ollmcp.json -m <model>

/tools inside ollmcp turns individual tools on and off. Model choice matters more here than with a hosted model: a model that is too small may call the wrong tool, invent arguments, or stop after one call. If that happens, try a larger one. That's a model problem, not a safety one. The server enforces the same rules whatever is calling it.

Then ask it things like "what's in app_db and how do orders relate to customers?" or "why is this query slow?".

Not in v0.1

Write mode, more than one server per process, HTTP transport, replication / InnoDB / lock diagnostics, Enterprise Edition features, EXPLAIN ANALYZE, stored functions.

Development

git clone https://github.com/pantoniades/sakila-mcp.git
cd sakila-mcp
uv sync
uv run ruff check . && uv run ruff format --check . && uv run mypy && uv run pytest

Integration tests need a MySQL server prepared once with tests/integration/setup.sh. Give it the address the server sees the tests connect from and a letters-and-digits password (extra arguments go to the mysql client), then point the tests at the server:

SAKILA_MCP_TEST_CLIENT_HOST=127.0.0.1 SAKILA_MCP_TEST_PASSWORD=<password> \
  tests/integration/setup.sh -u root -p
echo 'SAKILA_MCP_TEST_DSN=mysql://sakila_mcp_test:<password>@<host>:3306/sakila_mcp_test' > .env.test

.env.test is gitignored. Without it the integration tests skip.

One test is opt-in: it drives Claude Code (claude -p) through every tool against the test server and checks that the client accepts each result. It needs the claude CLI, uses your account, and never runs in CI:

uv run pytest -m llm
PYPI
sakila-mcp
Install Command
pip install sakila-mcp
Runtime: uvx