StarRocks Official MCP Server

The StarRocks MCP Server acts as a bridge between AI assistants and StarRocks databases. It allows for direct SQL execution, database exploration, data visualization via charts, and retrieving detailed schema/data overviews without requiring complex client-side setup.

Features

• Direct SQL Execution: Run SELECT queries (read_query) and DDL/DML commands (write_query).
• Database Exploration: List databases and tables, retrieve table schemas (starrocks:// resources).
• System Information: Access internal StarRocks metrics and states via the proc:// resource path.
• Detailed Overviews: Get comprehensive summaries of tables (table_overview) or entire databases (db_overview), including column definitions, row counts, and sample data.
• Data Visualization: Execute a query and generate a Plotly chart directly from the results (query_and_plotly_chart).
• Intelligent Caching: Table and database overviews are cached in memory to speed up repeated requests. Cache can be bypassed when needed.
• Flexible Configuration: Set connection details and behavior via environment variables.

Configuration

The MCP server is typically run via an MCP host. Configuration is passed to the host, specifying how to launch the StarRocks MCP server process.

Using Streamable HTTP (recommended):

To start the server in Streamable HTTP mode:

First test connect is ok:

$ STARROCKS_URL=root:@localhost:8000 uv run mcp-server-starrocks --test

Start the server:

uv run mcp-server-starrocks --mode streamable-http --port 8000

Then config the MCP like this:

{
  "mcpServers": {
    "mcp-server-starrocks": {
      "url": "http://localhost:8000/mcp"
    }
  }
}

Using uv with installed package (individual environment variables):

{
  "mcpServers": {
    "mcp-server-starrocks": {
      "command": "uv",
      "args": ["run", "--with", "mcp-server-starrocks", "mcp-server-starrocks"],
      "env": {
        "STARROCKS_HOST": "default localhost",
        "STARROCKS_PORT": "default 9030",
        "STARROCKS_USER": "default root",
        "STARROCKS_PASSWORD": "default empty",
        "STARROCKS_DB": "default empty"
      }
    }
  }
}

Using uv with installed package (connection URL):

{
  "mcpServers": {
    "mcp-server-starrocks": {
      "command": "uv",
      "args": ["run", "--with", "mcp-server-starrocks", "mcp-server-starrocks"],
      "env": {
        "STARROCKS_URL": "root:password@localhost:9030/my_database"
      }
    }
  }
}

Using uv with local directory (for development):

{
  "mcpServers": {
    "mcp-server-starrocks": {
      "command": "uv",
      "args": [
        "--directory",
        "path/to/mcp-server-starrocks", // <-- Update this path
        "run",
        "mcp-server-starrocks"
      ],
      "env": {
        "STARROCKS_HOST": "default localhost",
        "STARROCKS_PORT": "default 9030",
        "STARROCKS_USER": "default root",
        "STARROCKS_PASSWORD": "default empty",
        "STARROCKS_DB": "default empty"
      }
    }
  }
}

Using uv with local directory and connection URL:

{
  "mcpServers": {
    "mcp-server-starrocks": {
      "command": "uv",
      "args": [
        "--directory",
        "path/to/mcp-server-starrocks", // <-- Update this path
        "run",
        "mcp-server-starrocks"
      ],
      "env": {
        "STARROCKS_URL": "root:password@localhost:9030/my_database"
      }
    }
  }
}

Command-line Arguments:

The server supports the following command-line arguments:

uv run mcp-server-starrocks --help

• --mode {stdio,sse,http,streamable-http}: Transport mode (default: stdio or MCP_TRANSPORT_MODE env var)
• --host HOST: Server host for HTTP modes (default: localhost)
• --port PORT: Server port for HTTP modes
• --test: Run in test mode to verify functionality

Examples:

# Start in streamable HTTP mode on custom host/port
uv run mcp-server-starrocks --mode streamable-http --host 0.0.0.0 --port 8080

# Start in stdio mode (default)
uv run mcp-server-starrocks --mode stdio

# Run test mode
uv run mcp-server-starrocks --test

• The url field should point to the Streamable HTTP endpoint of your MCP server (adjust host/port as needed).
• With this configuration, clients can interact with the server using standard JSON over HTTP POST requests. No special SDK is required.
• All tool APIs accept and return standard JSON as described above.

Note: The sse (Server-Sent Events) mode is deprecated and no longer maintained. Please use Streamable HTTP mode for all new integrations.

Environment Variables:

Connection Configuration

You can configure StarRocks connection using either individual environment variables or a single connection URL:

Option 1: Individual Environment Variables

• STARROCKS_HOST: (Optional) Hostname or IP address of the StarRocks FE service. Defaults to localhost.
• STARROCKS_PORT: (Optional) MySQL protocol port of the StarRocks FE service. Defaults to 9030.
• STARROCKS_USER: (Optional) StarRocks username. Defaults to root.
• STARROCKS_PASSWORD: (Optional) StarRocks password. Defaults to empty string.
• STARROCKS_PASSWORD_KEYCHAIN_SERVICE: (Optional, macOS only) Generic password service name to use when reading the password from Keychain. This is only used when no explicit password is provided via STARROCKS_PASSWORD or STARROCKS_URL.
• STARROCKS_PASSWORD_KEYCHAIN_ACCOUNT: (Optional, macOS only) Generic password account name to use when reading the password from Keychain. Defaults to the resolved StarRocks user.
• STARROCKS_DB: (Optional) Default database to use if not specified in tool arguments or resource URIs. If set, the connection will attempt to USE this database. Tools like table_overview and db_overview will use this if the database part is omitted in their arguments. Defaults to empty (no default database).

Option 2: Connection URL (takes precedence over individual variables)

• STARROCKS_URL: (Optional) A connection URL string that contains all connection parameters in a single variable. Format: [<schema>://]user:password@host:port/database. The schema part is optional. When this variable is set, it takes precedence over the individual STARROCKS_HOST, STARROCKS_PORT, STARROCKS_USER, STARROCKS_PASSWORD, and STARROCKS_DB variables.

  Examples:

  • root:mypass@localhost:9030/test_db
  • mysql://admin:secret@db.example.com:9030/production
  • starrocks://user:pass@192.168.1.100:9030/analytics

Password precedence:

• A password embedded in STARROCKS_URL wins, including an explicit empty password like user:@host:9030/db.
• If STARROCKS_URL omits the password, STARROCKS_PASSWORD is used when set.
• If neither explicit password source is set and STARROCKS_PASSWORD_KEYCHAIN_SERVICE is configured, the password is read from macOS Keychain.

macOS Keychain example

Store the password:

security add-generic-password -U -a root -s mcp-server-starrocks -w 'secret'

Verify the stored password:

security find-generic-password -a root -s mcp-server-starrocks -w

Use it with this server:

export STARROCKS_URL=root@localhost:9030/test_db
export STARROCKS_PASSWORD_KEYCHAIN_SERVICE=mcp-server-starrocks
export STARROCKS_PASSWORD_KEYCHAIN_ACCOUNT=root

Additional Configuration

• STARROCKS_OVERVIEW_LIMIT: (Optional) An approximate character limit for the total text generated by overview tools (table_overview, db_overview) when fetching data to populate the cache. This helps prevent excessive memory usage for very large schemas or numerous tables. Defaults to 20000.

• STARROCKS_MCP_OUTPUT_DIR: (Optional) Directory used by read_query when its output_file argument is a relative path. Defaults to ~/.mcp-server-starrocks/output/. The directory is created on demand. Absolute paths passed to output_file (including ~-prefixed paths) bypass this setting. Note: files are written on the machine where the MCP server runs. For Claude Code / Claude Desktop the server runs locally, so files land on your laptop. For remote/http deployments the file lands on the server, not the client.

• STARROCKS_MYSQL_AUTH_PLUGIN: (Optional) Specifies the authentication plugin to use when connecting to the StarRocks FE service. For example, set to mysql_clear_password if your StarRocks deployment requires clear text password authentication (such as when using certain LDAP or external authentication setups). Only set this if your environment specifically requires it; otherwise, the default auth_plugin is used.

• MCP_TRANSPORT_MODE: (Optional) Communication mode that specifies how the MCP Server exposes its services. Available options:

  • stdio (default): Communicates through standard input/output, suitable for MCP Host hosting.
  • streamable-http (Streamable HTTP): Starts as a Streamable HTTP Server, supporting RESTful API calls.
  • sse: (Deprecated, not recommended) Starts in Server-Sent Events (SSE) streaming mode, suitable for scenarios requiring streaming responses. Note: SSE mode is no longer maintained, it is recommended to use Streamable HTTP mode uniformly.

Components

Tools

• read_query

  • Description: Execute a SELECT query or other commands that return a ResultSet (e.g., SHOW, DESCRIBE). Optionally write the full result to a local file instead of returning it inline — useful for results too large to fit in the model context.
  • Input:
    hljs language-json

    Copy

    {
      "query": "SQL query string",
      "db": "database name (optional, uses default database if not specified)",
      "output_file": "optional path; if set, writes the full result to disk and returns only a summary + small preview. Relative paths resolve against STARROCKS_MCP_OUTPUT_DIR (default: ~/.mcp-server-starrocks/output/); absolute paths and ~ are used as-is",
      "output_format": "optional: csv | tsv | json | jsonl. If omitted, inferred from output_file extension (.csv/.tsv/.json/.jsonl/.ndjson); defaults to csv"
    }
    

  • Output: Without output_file, text content containing the query results in CSV-like format with a header row and row count summary. With output_file, a short summary including the resolved absolute path, byte count, and row count, plus a small preview. Returns an error message on failure.

• write_query

  • Description: Execute a DDL (CREATE, ALTER, DROP), DML (INSERT, UPDATE, DELETE), or other StarRocks command that does not return a ResultSet.
  • Input:
    hljs language-json

    Copy

    {
      "query": "SQL command string",
      "db": "database name (optional, uses default database if not specified)"
    }
    

  • Output: Text content confirming success (e.g., "Query OK, X rows affected") or reporting an error. Changes are committed automatically on success.

• analyze_query

  • Description: Analyze a query and get analyze result using query profile or explain analyze.
  • Input:
    hljs language-json

    Copy

    {
      "uuid": "Query ID, a string composed of 32 hexadecimal digits formatted as 8-4-4-4-12",
      "sql": "Query SQL to analyze",
      "db": "database name (optional, uses default database if not specified)"
    }
    

  • Output: Text content containing the query analysis results. Uses ANALYZE PROFILE FROM if uuid is provided, otherwise uses EXPLAIN ANALYZE if sql is provided.

• query_and_plotly_chart

  • Description: Executes a SQL query, loads the results into a Pandas DataFrame, and generates a Plotly chart using a provided Python expression. Designed for visualization in supporting UIs.
  • Input:
    hljs language-json

    Copy

    {
      "query": "SQL query to fetch data",
      "plotly_expr": "Python expression string using 'px' (Plotly Express) and 'df' (DataFrame). Example: 'px.scatter(df, x=\"col1\", y=\"col2\")'",
      "db": "database name (optional, uses default database if not specified)"
    }
    

  • Output: A list containing:
    1. TextContent: A text representation of the DataFrame and a note that the chart is for UI display.
    2. ImageContent: The generated Plotly chart encoded as a base64 PNG image (image/png). Returns text error message on failure or if the query yields no data.

• table_overview

  • Description: Get an overview of a specific table: columns (from DESCRIBE), total row count, and sample rows (LIMIT 3). Uses an in-memory cache unless refresh is true.
  • Input:
    hljs language-json

    Copy

    {
      "table": "Table name, optionally prefixed with database name (e.g., 'db_name.table_name' or 'table_name'). If database is omitted, uses STARROCKS_DB environment variable if set.",
      "refresh": false // Optional, boolean. Set to true to bypass the cache. Defaults to false.
    }
    

  • Output: Text content containing the formatted overview (columns, row count, sample data) or an error message. Cached results include previous errors if applicable.

• db_overview

  • Description: Get an overview (columns, row count, sample rows) for all tables within a specified database. Uses the table-level cache for each table unless refresh is true.
  • Input:
    hljs language-json

    Copy

    {
      "db": "database_name", // Optional if default database is set.
      "refresh": false // Optional, boolean. Set to true to bypass the cache for all tables in the DB. Defaults to false.
    }
    

  • Output: Text content containing concatenated overviews for all tables found in the database, separated by headers. Returns an error message if the database cannot be accessed or contains no tables.

Resources

Direct Resources

• starrocks:///databases
  • Description: Lists all databases accessible to the configured user.
  • Equivalent Query: SHOW DATABASES
  • MIME Type: text/plain

Resource Templates

• starrocks:///{db}/{table}/schema

  • Description: Gets the schema definition of a specific table.
  • Equivalent Query: SHOW CREATE TABLE {db}.{table}
  • MIME Type: text/plain

• starrocks:///{db}/tables

  • Description: Lists all tables within a specific database.
  • Equivalent Query: SHOW TABLES FROM {db}
  • MIME Type: text/plain

• proc:///{+path}

  • Description: Accesses StarRocks internal system information, similar to Linux /proc. The path parameter specifies the desired information node.
  • Equivalent Query: SHOW PROC '/{path}'
  • MIME Type: text/plain
  • Common Paths:
    • /frontends - Information about FE nodes.
    • /backends - Information about BE nodes (for non-cloud native deployments).
    • /compute_nodes - Information about CN nodes (for cloud native deployments).
    • /dbs - Information about databases.
    • /dbs/<DB_ID> - Information about a specific database by ID.
    • /dbs/<DB_ID>/<TABLE_ID> - Information about a specific table by ID.
    • /dbs/<DB_ID>/<TABLE_ID>/partitions - Partition information for a table.
    • /transactions - Transaction information grouped by database.
    • /transactions/<DB_ID> - Transaction information for a specific database ID.
    • /transactions/<DB_ID>/running - Running transactions for a database ID.
    • /transactions/<DB_ID>/finished - Finished transactions for a database ID.
    • /jobs - Information about asynchronous jobs (Schema Change, Rollup, etc.).
    • /statistic - Statistics for each database.
    • /tasks - Information about agent tasks.
    • /cluster_balance - Load balance status information.
    • /routine_loads - Information about Routine Load jobs.
    • /colocation_group - Information about Colocation Join groups.
    • /catalog - Information about configured catalogs (e.g., Hive, Iceberg).

Prompts

None defined by this server.

Caching Behavior

• The table_overview and db_overview tools utilize an in-memory cache to store the generated overview text.
• The cache key is a tuple of (database_name, table_name).
• When table_overview is called, it checks the cache first. If a result exists and the refresh parameter is false (default), the cached result is returned immediately. Otherwise, it fetches the data from StarRocks, stores it in the cache, and then returns it.
• When db_overview is called, it lists all tables in the database and then attempts to retrieve the overview for each table using the same caching logic as table_overview (checking cache first, fetching if needed and refresh is false or cache miss). If refresh is true for db_overview, it forces a refresh for all tables in that database.
• The STARROCKS_OVERVIEW_LIMIT environment variable provides a soft target for the maximum length of the overview string generated per table when populating the cache, helping to manage memory usage.
• Cached results, including any error messages encountered during the original fetch, are stored and returned on subsequent cache hits.

Debug

After starting mcp server, you can use inspector to debug:

npx @modelcontextprotocol/inspector

Demo