🦆 DuckDB MCP Server

Give your AI assistant a blazing-fast, persistent SQL engine.

A minimal Model Context Protocol server that plugs DuckDB straight into Claude — query CSVs, Parquet, and cloud data in plain language, at in-process speed.

⚡ ~10 ms / query · 🪶 2 runtime deps · 🧰 12 tools · 🔒 read-only & secret-safe

✨ Why this one?

Related MCP server: mcp-server-motherduck

🚀 Quick start

1. Run it — no install needed (via uv):

uvx --from git+https://github.com/wuqunfei/duckdb-mcp-mini duckdb-mcp --db :memory:

2. Point Claude Desktop at it — add this to claude_desktop_config.json and restart Claude:

{
  "mcpServers": {
    "duckdb": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/wuqunfei/duckdb-mcp-mini", "duckdb-mcp", "--db", ":memory:"]
    }
  }
}

3. Ask away 💬

"Load ~/data/sales.csv and show me total revenue by region."

That's it. Tools are available immediately. 🎉

📦 Install

pip install .            # from a clone, into the current environment
pip install -e ".[dev]"  # for development (editable + dev tools)

Or run straight from Git without installing:

uvx --from git+https://github.com/wuqunfei/duckdb-mcp-mini duckdb-mcp --db :memory:

ℹ️ This package is not published to PyPI — install from source or run from Git as shown above.

Requirements: 🐍 Python 3.11–3.14 · 🦆 DuckDB 1.5.2 (pinned in pyproject.toml, easy to change) · 🔌 MCP SDK 2.0.0+

🏃 Run

The installed console script is duckdb-mcp. It speaks MCP over stdio, so you'll normally launch it from an MCP client — but you can start it directly too:

duckdb-mcp                                             # in-memory, read-write
duckdb-mcp --db /path/to/analytics.duckdb --schema main
duckdb-mcp --db analytics.duckdb --init-sql init.sql
duckdb-mcp --db analytics.duckdb --read-only

python -m duckdb_mcp.cli --db :memory:                 # module form
uv run duckdb-mcp --db :memory:                        # via uv, no install

CLI arguments

🐛 Debugging: run with --log-level DEBUG (or LOG_LEVEL=DEBUG) to log every incoming tool request — name and arguments — to stderr, so you can see exactly what a client sends:

DEBUG duckdb_mcp: tool request: query args={'sql': 'SELECT 42 AS answer'}

Logs always go to stderr (never stdout, which is the protocol channel in stdio mode).

🔒 Read-only mode

--read-only opens the connection via DuckDB's own read-only flag, so writes are blocked at the engine level (not by inspecting SQL):

• ✅ SELECT works normally

• 🚫 INSERT / UPDATE / DELETE / CREATE / DROP are rejected by DuckDB

• 💡 Perfect for shared analytics/reporting databases where accidental writes must be impossible

🔌 Transports

The server speaks two MCP transports — pick with --transport:

# stdio (default) — the process talks over stdin/stdout
duckdb-mcp --db analytics.duckdb

🌐 Run an HTTP server

Configuration splits cleanly in two: args set the database and network binding, environment variables carry secrets and toggles (so they stay out of ps and shell history).

# 1) Environment — secrets & toggles
export MCP_AUTH_TOKEN="a-long-random-secret"     # require Bearer auth (recommended)
export ALLOW_UNSIGNED_EXTENSIONS=true            # optional: allow community extensions
export AWS_ACCESS_KEY_ID="${AWS_ACCESS_KEY_ID}"  # optional: for s3:// queries
export AWS_SECRET_ACCESS_KEY="${AWS_SECRET_ACCESS_KEY}"

# 2) Args — database + network binding
duckdb-mcp \
  --transport http \
  --host 0.0.0.0 \
  --port 8000 \
  --db /data/analytics.duckdb \
  --init-sql /data/init.sql \
  --read-only

# → serving at  http://0.0.0.0:8000/mcp

--init-sql runs a SQL file once, when the connection opens. Combined with --read-only it must be read-only-safe (loading extensions or setting options is fine — creating tables is not):

-- /data/init.sql
INSTALL httpfs; LOAD httpfs;     -- read remote/S3 data
SET memory_limit = '4GB';

Drop --read-only if your init script needs to create tables or views.

No install? Run the same thing straight from Git:

MCP_AUTH_TOKEN="a-long-random-secret" \
uvx --from git+https://github.com/wuqunfei/duckdb-mcp-mini duckdb-mcp \
  --transport http --host 0.0.0.0 --port 8000 --db :memory:

Verify it's up (with MCP_AUTH_TOKEN set, a missing/wrong token returns 401):

curl -s -o /dev/null -w '%{http_code}\n' \
  -X POST http://127.0.0.1:8000/mcp \
  -H 'Authorization: Bearer a-long-random-secret' \
  -H 'Accept: application/json, text/event-stream' \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"ping"}'
# 401 = missing/invalid token   ·   400 = reached the server (a real client does the full MCP handshake)

🔑 Bearer token

Setting MCP_AUTH_TOKEN turns on auth for the HTTP transport:

• 🌱 Environment only — read from MCP_AUTH_TOKEN, never a CLI flag, so it stays out of ps and shell history.

• 🔌 HTTP only — ignored for stdio (the client owns the process); setting it there prints a warning.

• ⚠️ Access control, not identity — every caller with the token gets full database access. Pair it with --read-only and network controls; it is not per-user auth.

🔒 --host defaults to 127.0.0.1 (localhost only). Bind to 0.0.0.0 only on a trusted network, and always with MCP_AUTH_TOKEN set.

🔗 Connect a client

MCP client config (Claude Desktop or any client that supports an HTTP URL) — omit headers if you didn't set a token:

{
  "mcpServers": {
    "duckdb": {
      "url": "http://127.0.0.1:8000/mcp",
      "headers": { "Authorization": "Bearer a-long-random-secret" }
    }
  }
}

From Python (the mcp client libraries ship httpx2):

import asyncio
import httpx2
from mcp import ClientSession
from mcp.client.streamable_http import streamable_http_client

URL = "http://127.0.0.1:8000/mcp"
HEADERS = {"Authorization": "Bearer a-long-random-secret"}  # omit if no MCP_AUTH_TOKEN

async def main():
    async with httpx2.AsyncClient(headers=HEADERS) as http:
        async with streamable_http_client(URL, http_client=http) as streams:
            read, write = streams[0], streams[1]
            async with ClientSession(read, write) as session:
                await session.initialize()
                tools = await session.list_tools()
                print("tools:", [t.name for t in tools.tools])
                result = await session.call_tool("query", {"sql": "SELECT 42 AS answer"})
                print(result.content[0].text)

asyncio.run(main())

🖥️ Configure Claude Desktop

Add one of the following to your claude_desktop_config.json, then restart Claude.

Run from Git (no install):

{
  "mcpServers": {
    "duckdb": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/wuqunfei/duckdb-mcp-mini", "duckdb-mcp", "--db", ":memory:"]
    }
  }
}

Run from a local clone:

{
  "mcpServers": {
    "duckdb": {
      "command": "uv",
      "args": ["run", "--directory", "/absolute/path/to/duckdb-mcp-mini", "duckdb-mcp", "--db", "/home/user/analytics.duckdb", "--schema", "main"]
    }
  }
}

After pip install . (command on PATH):

{
  "mcpServers": {
    "duckdb": {
      "command": "duckdb-mcp",
      "args": ["--db", "/home/user/analytics.duckdb", "--read-only"]
    }
  }
}

📍 Config file locations

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Linux: ~/.config/Claude/claude_desktop_config.json

• Windows: %APPDATA%\Claude\claude_desktop_config.json

🔁 Multiple servers

Add more entries under mcpServers with distinct names (e.g. duckdb-dev, duckdb-prod), each with its own --db/--schema.

☁️ Environment variables & cloud data

Any variables set in the client's env block are available to DuckDB during connection (handy for S3/cloud credentials). Values support ${VAR_NAME} interpolation, so you can reference the system environment instead of hardcoding secrets into the config file:

{
  "mcpServers": {
    "duckdb": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/wuqunfei/duckdb-mcp-mini", "duckdb-mcp", "--db", ":memory:"],
      "env": {
        "AWS_ACCESS_KEY_ID": "${AWS_ACCESS_KEY_ID}",
        "AWS_SECRET_ACCESS_KEY": "${AWS_SECRET_ACCESS_KEY}",
        "AWS_REGION": "${AWS_REGION}",
        "S3_BUCKET": "my-data-bucket"
      }
    }
  }
}

Set the referenced variables in your shell first:

export AWS_ACCESS_KEY_ID="your-key"
export AWS_SECRET_ACCESS_KEY="your-secret"
export AWS_REGION="us-east-1"

Then query cloud data directly:

SELECT * FROM read_parquet('s3://my-data-bucket/data.parquet')

🔐 Interpolation happens before the DuckDB connection is opened, so credentials are ready in time for connection setup and any --init-sql. Use the list_environments tool to confirm what's set — values are always masked.

🧩 Community & unsigned extensions

DuckDB only loads signed extensions by default. To use community or self-hosted extensions, enable unsigned extensions with the --allow-unsigned-extensions flag or the ALLOW_UNSIGNED_EXTENSIONS environment variable (default: off):

duckdb-mcp --db :memory: --allow-unsigned-extensions
# or
ALLOW_UNSIGNED_EXTENSIONS=true duckdb-mcp --db :memory:

In a Claude Desktop config:

{
  "mcpServers": {
    "duckdb": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/wuqunfei/duckdb-mcp-mini", "duckdb-mcp", "--db", ":memory:", "--allow-unsigned-extensions"]
    }
  }
}

The flag is applied as a connection-time setting, so it's already on — just INSTALL and LOAD through the execute tool. For example, neuesql/atm_talib (TA-Lib for DuckDB):

INSTALL talib FROM 'https://neuesql.github.io/atm_talib';
LOAD talib;

⚠️ Unsigned extensions run native code with full trust. Only enable this for extensions from sources you trust.

🧰 Tools (12)

query vs execute — query is for reads (fetches rows, formats a table); execute is for writes/DDL (runs the statement, returns "Executed successfully").

🏗️ Architecture

launch (cli.main)
   ↓
DuckDBSession(...)                     # src/duckdb_mcp/session.py
   ├─ interpolate ${VAR} across os.environ   ← runs BEFORE connecting
   ├─ duckdb.connect(db, read_only=...)      ← one persistent connection
   └─ run --init-sql (if provided)           ← falls back to :memory: on error
   ↓
create_server(session)                 # src/duckdb_mcp/server.py
   └─ registers 12 tools on an mcp MCPServer; each calls dispatch_tool()
   ↓
_serve(session, transport=...)         # src/duckdb_mcp/cli.py
   └─ run_stdio_async() | run_streamable_http_async()

• 🧩 session.py holds DuckDBSession with no MCP imports, so the connection/formatting logic is unit-testable with only duckdb.

• 🗂️ server.py holds the _HANDLERS registry (single source of truth for the tool set) plus dispatch_tool() and the thin typed MCP registrations.

• 🎛️ cli.py parses args, builds the session, and serves.

Good to know:

• 🛟 If connecting to --db (or running --init-sql) fails, the session degrades to an in-memory read-write connection rather than crashing.

• ⚠️ SQL is f-string interpolated (table names, filepaths). This is intentional for a local single-user tool — inputs are not sanitized.

🛠️ Development

pip install -e ".[dev]"      # or: uv sync --extra dev

black --check src tests      # format (drop --check to apply)
ruff check src tests         # lint
mypy src                     # type check
pytest                       # tests

pytest tests/test_server.py::test_dispatch_query   # run a single test

📁 Project layout

duckdb-mcp-mini/
├── src/duckdb_mcp/
│   ├── __init__.py        # package version + DuckDBSession export
│   ├── session.py         # persistent DuckDB session (no MCP deps)
│   ├── server.py          # tool registry + dispatch + MCP server wiring
│   └── cli.py             # argument parsing + entrypoint
├── tests/                 # pytest suite
├── .github/workflows/
│   ├── ci.yml             # lint + type + test on py3.11–3.14
│   └── release.yml        # build on tag (PyPI publish disabled)
├── pyproject.toml
├── LICENSE
└── README.md

🔄 CI / CD

• CI (.github/workflows/ci.yml) runs black, ruff, mypy, and pytest on Python 3.11, 3.12, 3.13, and 3.14 for every push/PR to main, and checks that duckdb-mcp --help works.

• CD (.github/workflows/release.yml) builds the sdist/wheel on a v* tag and attaches them to the GitHub Release. PyPI publishing is intentionally disabled — the publish-pypi job is gated behind if: false; see the comments in that file to enable it later.

🤝 Contributing

Issues and PRs are welcome! Please keep the footprint small (the minimalism is a feature 🪶), and make sure black / ruff / mypy / pytest all pass before opening a PR.

Built with 🦆 + 🔌 · Licensed under MIT