cubrid-mcp-server

cubrid-mcp-server

A Model Context Protocol server for CUBRID, enabling LLMs to safely inspect schemas and execute read-only queries via pycubrid.

Terminal recording of a real MCP stdio session (demos/mcp_demo.py, recorded with demos/record_demo.sh).

Features

Tool	Description
all_table_names	List every user table in the database
filter_table_names	Substring search over table names
schema_definitions	Column types, nullability, defaults, and primary key info
describe_table	Full metadata: columns, primary key, and indexes in one call
list_indexes	Indexes for a table with key columns and flags
explain_query	Execution plan/trace for a SELECT/WITH (via CUBRID SHOW TRACE)
table_row_counts	COUNT(*) for one or many tables
list_serials	CUBRID SERIAL sequences with current value and bounds
list_class_hierarchy	CUBRID CLASS inheritance relationships
execute_query	Run read-only SQL with automatic output truncation
health_check	Verify database connectivity on demand
execute_write	Run a single INSERT/UPDATE/DELETE in an atomic transaction (only registered when opt-in write mode is enabled)

Resources

Schema metadata is also exposed as read-only MCP Resources, so clients can discover and read schema context without a tool call. Resources reuse the same read-only catalog queries as the tools — no additional data access or write surface.

Resource URI	MIME type	Description
cubrid://agent-guide	text/markdown	CUBRID agent guide: dialect, types, performance
cubrid://guide/*	text/markdown	4 topic guides: sql-dialect, types, performance, collections
cubrid://schema	application/json	Whole-schema index: every user table with its per-table resource URI
cubrid://schema/{table}	application/json	Per-table metadata (columns, primary key, indexes) — mirrors describe_table

Schema resources always read the default connection (see Multi-Connection). The guides are static Markdown. Table names in {table} are percent-decoded by URI-template matching; an unknown or system table produces a resource-read error, matching the describe_table tool.

Prompts

The server also exposes a small set of MCP Prompt templates — reusable, guided starting points for common inspection tasks. Prompts are guidance-only: each one returns text that tells the client which of the existing read-only tools to call and in what order. They never touch the database, execute SQL, or add any new data-access surface, and any argument you pass is fenced and treated strictly as untrusted data. The prompts are advisory templates only — the actual read-only enforcement remains in the underlying tools (execute_query/explain_query via safety.py).

Prompt	Arguments	Description
summarize_table	table	Describe a table, then sample it with a bounded read-only query
explain_query	sql	Obtain and interpret a SELECT/WITH execution plan via explain_query
inspect_schema	(none)	Build a high-level overview of the whole schema from the read-only tools
find_index_candidates	table	Review a table's index coverage for potential review areas
optimize_query	sql	Analyze the execution plan and suggest CUBRID-specific optimizations
migrate_from_mysql	sql	Convert MySQL query syntax to valid CUBRID SQL
explore_unknown_db	(none)	Systematically explore an unfamiliar database
safe_data_analysis	question	Answer data questions using read-only queries
write_cubrid_sql	natural_language	Generate valid CUBRID SQL from natural language

Quick Start

Configure

Set the required environment variables:

export CUBRID_HOST=localhost
export CUBRID_PORT=33000        # optional, default: 33000
export CUBRID_USER=readonly_user   # a CUBRID user with SELECT-only grants (see Security)
export CUBRID_PASSWORD=secret
export CUBRID_DATABASE=mydb

Optional settings:

Variable	Default	Description
CUBRID_MCP_READONLY	1	Enforce read-only SQL whitelist. 0 relaxes the whitelist only; execute_query still rejects write, DDL, transaction-control, CALL, EVALUATE and KILL statements (use execute_write)
CUBRID_MCP_MAX_CHARS	4000	Max characters in query output
CUBRID_MCP_MAX_ROWS	1000	Max rows returned by execute_query before truncation
CUBRID_MCP_MAX_SQL_LENGTH	65536	Max length (characters) of a submitted SQL statement
CUBRID_MCP_QUERY_TIMEOUT	30	Per-statement socket read timeout in seconds. If the server sends no data within this window the query is aborted and the connection is reset. This is a socket read timeout, not a true server-side statement timeout.
CUBRID_MCP_AUDIT_LOG	0	Opt-in audit logging. When enabled, emits one redaction-safe JSON record per executed statement (execute_query/explain_query/execute_write) to stderr — see Logging. Honoured per connection (CUBRID_<NAME>_MCP_AUDIT_LOG).
CUBRID_MCP_WRITE	0	Opt-in write mode. When enabled (1), registers the execute_write tool for single-statement DML. Honoured per connection (CUBRID_<NAME>_MCP_WRITE); the tool is registered when any connection enables it. Off by default.

Multiple connections

By default the bare CUBRID_* variables define a single connection named default. You can serve additional CUBRID databases from the same process by listing extra connection names in CUBRID_CONNECTIONS (comma-separated) and providing CUBRID_<NAME>_* variables for each. Every tool accepts an optional connection argument selecting which connection to target; omitting it (or passing default) uses the bare-variable connection, so existing single-database setups are unchanged.

# Default connection (unchanged)
export CUBRID_HOST=localhost
export CUBRID_USER=readonly_user
export CUBRID_PASSWORD=secret
export CUBRID_DATABASE=mydb

# Additional named connections
export CUBRID_CONNECTIONS=reporting,analytics

export CUBRID_REPORTING_HOST=reporting-db
export CUBRID_REPORTING_USER=readonly_user
export CUBRID_REPORTING_PASSWORD=secret
export CUBRID_REPORTING_DATABASE=reports
export CUBRID_REPORTING_MCP_MAX_ROWS=500   # optional per-connection tuning

export CUBRID_ANALYTICS_HOST=analytics-db
export CUBRID_ANALYTICS_USER=readonly_user
export CUBRID_ANALYTICS_PASSWORD=secret
export CUBRID_ANALYTICS_DATABASE=analytics

Notes:

• Connection names must match [A-Za-z0-9_]+ and are matched case-insensitively.
• default is reserved (it always comes from the bare CUBRID_* variables) and cannot appear in CUBRID_CONNECTIONS.
• For a named connection <NAME>, connection fields live at CUBRID_<NAME>_HOST etc. and the optional MCP tuning knobs at CUBRID_<NAME>_MCP_* (same suffixes as the global ones). Named connections do not inherit values from the bare vars.
• Selecting an unknown connection returns a clear error listing the available names.
• Each connection has its own read-only enforcement, so a named connection can set CUBRID_<NAME>_MCP_READONLY independently of the default.

Run

Run from PyPI

Use uvx to run directly from PyPI:

uvx cubrid-mcp-server

Or with pipx:

pipx run cubrid-mcp-server

Run from source

git clone https://github.com/cubrid-lab/cubrid-mcp-server.git
cd cubrid-mcp-server
python -m venv .venv && source .venv/bin/activate
pip install -e .
cubrid-mcp-server

MCP Client Integration

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "cubrid": {
      "command": "uvx",
      "args": ["cubrid-mcp-server"],
      "env": {
        "CUBRID_HOST": "localhost",
        "CUBRID_USER": "readonly_user",
        "CUBRID_PASSWORD": "secret",
        "CUBRID_DATABASE": "mydb"
      }
    }
  }
}

Claude Code

Add to .mcp.json in your project root:

{
  "mcpServers": {
    "cubrid": {
      "command": "uvx",
      "args": ["cubrid-mcp-server"],
      "env": {
        "CUBRID_HOST": "localhost",
        "CUBRID_USER": "readonly_user",
        "CUBRID_PASSWORD": "secret",
        "CUBRID_DATABASE": "mydb"
      }
    }
  }
}

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "cubrid": {
      "command": "uvx",
      "args": ["cubrid-mcp-server"],
      "env": {
        "CUBRID_HOST": "localhost",
        "CUBRID_USER": "readonly_user",
        "CUBRID_PASSWORD": "secret",
        "CUBRID_DATABASE": "mydb"
      }
    }
  }
}

Security

The server is read-only by default. A code-level SQL whitelist allows only SELECT, SHOW, DESC, DESCRIBE, and WITH statements. Multi-statement queries are rejected. CUBRID has no EXPLAIN statement; use the explain_query tool for execution plans (it executes the statement under SET TRACE ON and rolls back).

Read-only mode also rejects INCR, DECR, serial advances (NEXT_VALUE, NEXTVAL, SERIAL_NEXT_VALUE) and the click counter WITH INCREMENT FOR: they change data inside a SELECT, the rollback does not undo them, and a SELECT-only grant does not prevent them. See the Security Model.

The SQL keyword checks are a guardrail and defense-in-depth, not a security boundary. The primary control is a dedicated least-privilege database account: run the server as a CUBRID user that has SELECT only on the tables the model needs, never reuse dba or the owner account, and keep the account to the minimum privileges required, with no DDL rights. See SECURITY.md.

For production use, also configure a read-only database user. See SECURITY.md for the recommended setup.

Write mode (opt-in)

Write access is disabled by default. Setting CUBRID_MCP_WRITE=1 registers an additional execute_write tool that accepts a single INSERT, UPDATE, or DELETE statement and runs it in an explicit transaction (commit on success, rollback on any error). When write mode is off the tool is not registered at all, so no write path is exposed in MCP capability discovery. With multiple connections configured, the tool is registered when any connection enables writes (via CUBRID_MCP_WRITE or CUBRID_<NAME>_MCP_WRITE).

Constraints and rationale:

• Single-statement DML only. Standalone reads, DDL (CREATE/ALTER/DROP/TRUNCATE), transaction-control, and multi-statement input are rejected. (A single DML statement may still legally contain subqueries, e.g. INSERT ... SELECT.)
• DDL is intentionally unsupported. execute_write is a DML tool (INSERT/UPDATE/DELETE) and execute_query is read-only, so CREATE/ALTER/DROP/TRUNCATE are rejected by both. (The server runs pycubrid with autocommit off; on the CUBRID versions tested (10.2, 11.2, 11.4) DDL is transactional and is rolled back if it is not committed, so the exclusion rests on the tool contracts, not on DDL auto-commit.)
• Write mode is per-connection. execute_write accepts the same optional connection argument as the read tools and runs against that connection; a connection whose CUBRID_<NAME>_MCP_WRITE is off refuses the write even when another connection enables it.
• execute_query never becomes a write path through the write-mode flag or CUBRID_MCP_READONLY: with CUBRID_MCP_READONLY=0 it still rejects write, DDL and transaction-control statements, method calls (CALL, EVALUATE) and KILL before they run, so CALL login(...) cannot change the session user, rejects INCR, DECR, serial advances and the click counter, which change data inside a SELECT, and rolls back any statement that returns no result set. Clients that sent writes through execute_query with CUBRID_MCP_READONLY=0 must switch to execute_write.
• Enforcement is defense-in-depth; still run the server as a CUBRID user granted only the privileges it needs. See SECURITY.md.

Logging

The server speaks the MCP stdio transport, where stdout carries the JSON-RPC protocol stream. Anything written to stdout by the server or its dependencies will corrupt that stream and break the client connection. For this reason all logging is routed to stderr, and you should keep it that way: when adding custom logging or diagnostics, never print() to stdout — use the standard logging module (which is configured to emit on stderr) or write to stderr explicitly. The log level defaults to INFO.

Errors surfaced back to the LLM client are sanitized: only the exception category (e.g. query failed: OperationalError) is returned, while the full exception detail is logged to stderr for operators. This keeps schema details, hostnames, SQL fragments, and configuration values out of client-visible messages.

Audit logging (opt-in)

Set CUBRID_MCP_AUDIT_LOG=1 to record every executed statement (execute_query, explain_query, and execute_write) as a single structured JSON line on stderr. It is off by default and honoured per connection — a named connection sets CUBRID_<NAME>_MCP_AUDIT_LOG independently of the default. Each record is redaction-safe and contains only:

Field	Description
tool	The MCP tool that ran the statement (execute_query / explain_query / execute_write).
status	ok or error.
category	The leading SQL keyword only (e.g. SELECT, WITH), read after comments and string literals are stripped — never the full statement.
identifiers	Table names extracted after FROM/JOIN via a strict identifier regex; anything that is not a bare identifier (values, literals, expressions) is dropped.
sql_length	Length of the submitted SQL, in characters.
row_count, truncated	Result size and whether output was truncated (success only).
duration_ms	Wall-clock duration of the call, in integer milliseconds.
error_type	On failure, the exception class name only (via the same sanitization as client-facing errors).

The raw SQL text, bound parameters, and literal values are never logged, so secrets embedded in a query (e.g. WHERE token = '...') do not reach the audit stream. Records go to stderr only and never to stdout.

Development

git clone https://github.com/cubrid-lab/cubrid-mcp-server.git
cd cubrid-mcp-server
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

# Lint & type check
ruff check .
mypy cubrid_mcp_server

# Unit tests
pytest -m "not integration"

# Integration tests (requires running CUBRID)
export CUBRID_HOST=localhost CUBRID_USER=dba CUBRID_PASSWORD="" CUBRID_DATABASE=demodb
pytest -m integration

The live shared-session concurrency regression can also be run on its own:

pytest tests/test_integration.py::TestCubridIntegration::test_concurrent_tool_calls_serialize_shared_session -q

It starts simultaneous in-process explain_query and execute_query calls on the same cached Database connection: the query must wait until the real trace context releases the existing RLock. It verifies independent responses, session reuse and cursor cleanup. This is not a test of MCP stdio transport concurrency, connection pooling or per-request transaction isolation.

Related Projects

• pycubrid — Pure-Python DB-API 2.0 driver this server is built on
• sqlalchemy-cubrid — SQLAlchemy 2.0–2.2 dialect for CUBRID
• cubrid-cookbook-python — 68 runnable examples incl. one-command app templates that use this stack

Disclaimer

This project is part of CUBRID Lab, an independent open-source initiative for CUBRID developer tooling, and is not affiliated with, sponsored by, or endorsed by CUBRID Corporation or the official CUBRID project.

License

MIT (see LICENSE).