Cloudera Hive Server

A standard Model Context Protocol server that exposes a Cloudera Data Warehouse Virtual Warehouse (Hive) to any MCP client — Claude Desktop, Claude Code, the Claude Agent SDK, LangChain, LlamaIndex, Cline, Continue, etc.

Tools

Tool	Purpose
list_databases	List every database in the Virtual Warehouse.
list_tables	List tables in a given database.
describe_table	Return columns, types, and comments for a table.
get_table_sample	Preview the first N rows (1–100) of a table.
execute_query	Run a HiveQL query. Read-only by default; row-capped for safety.

Prerequisites

• Python 3.10+
• Network access to your Cloudera Virtual Warehouse (typically *.dw.cloudera.site:443)
• A workload user + password with query privileges on the target VW

Recommended install: uvx from Git (zero-setup for clients)

uvx is the Python analog of npx. It fetches the package, resolves its dependencies into an isolated cache, and runs the entry point — no clone, no venv, no pip install on the client machine.

One-time on the client machine — install uv (ships uvx):

curl -LsSf https://astral.sh/uv/install.sh | sh    # macOS / Linux
# Windows:  powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

Then any MCP client can launch this server with:

uvx --from git+https://github.com/mjain/hive-mcp-server hive-mcp-server

(replace the Git URL with your fork / internal mirror)

Environment variables

The server needs Hive credentials in its environment. Every MCP client config below sets them via an env block — no .env file needed on the client.

Variable	Default	Description
HIVE_HOST	(required)	Virtual Warehouse hostname
HIVE_PORT	443	HTTPS port
HIVE_HTTP_PATH	/cliservice	HiveServer2 HTTP path
HIVE_USERNAME	(required)	Cloudera workload user
HIVE_PASSWORD	(required)	Cloudera workload password
HIVE_READ_ONLY	true	If true, execute_query rejects DDL/DML
HIVE_QUERY_ROW_LIMIT	1000	Max rows returned by execute_query
MCP_TRANSPORT	stdio	Transport passed to mcp.run(). Use stdio for MCP clients; set to sse when driving via the MCP Inspector.

---

Client setup

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "hive": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/mjain/hive-mcp-server",
        "hive-mcp-server"
      ],
      "env": {
        "HIVE_HOST": "your-vw-host.dw.cloudera.site",
        "HIVE_USERNAME": "your-workload-user",
        "HIVE_PASSWORD": "your-workload-password",
        "HIVE_READ_ONLY": "true"
      }
    }
  }
}

Fully quit Claude Desktop (⌘Q) and reopen. The five Hive tools appear in the tool tray.

Claude Code CLI

claude mcp add hive \
  --env HIVE_HOST=your-vw-host.dw.cloudera.site \
  --env HIVE_USERNAME=your-workload-user \
  --env HIVE_PASSWORD=your-workload-password \
  -- uvx --from git+https://github.com/mjain/hive-mcp-server hive-mcp-server

Add -s user before hive to make it available in every project.

Claude Agent SDK (Python)

# pip install claude-agent-sdk
import asyncio
from claude_agent_sdk import query, ClaudeAgentOptions

async def main():
    options = ClaudeAgentOptions(
        mcp_servers={
            "hive": {
                "type": "stdio",
                "command": "uvx",
                "args": [
                    "--from",
                    "git+https://github.com/mjain/hive-mcp-server",
                    "hive-mcp-server",
                ],
                "env": {
                    "HIVE_HOST": "your-vw-host.dw.cloudera.site",
                    "HIVE_USERNAME": "your-workload-user",
                    "HIVE_PASSWORD": "your-workload-password",
                    "HIVE_READ_ONLY": "true",
                },
            }
        },
        allowed_tools=[
            "mcp__hive__list_databases",
            "mcp__hive__list_tables",
            "mcp__hive__describe_table",
            "mcp__hive__get_table_sample",
            "mcp__hive__execute_query",
        ],
    )
    async for msg in query(
        prompt="List all Hive databases, then describe the largest table in the first one.",
        options=options,
    ):
        print(msg)

asyncio.run(main())

The same {command, args, env} shape works for LangChain's MCP adapter, LlamaIndex, Cline, Continue, Zed, and every other MCP client.

---

Local development

Clone and install in editable mode when working on the server itself:

git clone <this repo>
cd hive-mcp-server-claude
python -m venv .venv
source .venv/bin/activate
pip install -e .
cp .env.example .env    # fill in credentials
hive-mcp-server         # runs on stdio; Ctrl-C to stop

Smoke-test the connection:

python -c "from hive_mcp_server.tools.hive_tools import list_databases; print(list_databases())"

Code layout

The server follows the same pattern as cloudera/iceberg-mcp-server:

src/hive_mcp_server/
├── __init__.py           # version + main/mcp re-exports
├── server.py             # thin MCP registration layer (@mcp.tool wrappers)
└── tools/
    ├── __init__.py
    └── hive_tools.py     # config, connection, SQL safety, tool logic

To add a new tool: implement it in tools/hive_tools.py, then add a matching @mcp.tool() wrapper in server.py whose docstring becomes the tool description exposed to the LLM.

Transport

By default the server runs on stdio, which is what all MCP clients (Claude Desktop, Claude Code, etc.) expect. To use the MCP Inspector web UI instead:

MCP_TRANSPORT=sse hive-mcp-server

Publishing your own fork

Push to any Git host — clients reference the URL:

git init && git add . && git commit -m "initial"
git remote add origin https://github.com/<you>/hive-mcp-server
git push -u origin main

Then everyone points their uvx --from git+... at your URL. To publish to PyPI so clients can just say uvx hive-mcp-server (no --from):

pip install build twine
python -m build
twine upload dist/*

Safety

• HIVE_READ_ONLY=true (default) — execute_query rejects INSERT / UPDATE / DELETE / DROP / ALTER / TRUNCATE / CREATE / REPLACE / MERGE / GRANT / REVOKE / MSCK / LOAD / EXPORT / IMPORT.
• HIVE_QUERY_ROW_LIMIT=1000 — caps execute_query results so a SELECT * on a billion-row table doesn't blow up the agent's context.
• Identifier validation — database and table arguments must match [A-Za-z_][A-Za-z0-9_]*; anything else is rejected before reaching Hive.
• No credentials in tool arguments — the connection is configured entirely through environment variables; agents cannot see or override them.

License

MIT.