toolkit-mcp — an MCP server for Claude

A small but real Model Context Protocol server that gives Claude three capabilities it doesn't have on its own: querying a local SQLite database, searching files inside one directory, and looking up weather.

Built on the official Python SDK (mcp v2). Works with Claude Desktop, Claude Code, and any other MCP client.

The problem

Claude can't see your database, your files, or anything that happened after its training cutoff. Copy-pasting a schema and some rows into the chat works once and doesn't scale — and it gives Claude a stale snapshot rather than the ability to go and look.

Related MCP server: SQLite MCP Server

The solution

MCP is the standard interface for handing an assistant real tools. This server implements five of them, with the boring parts done properly: the database connection is genuinely read-only, file access can't escape its root, and every error comes back as data the model can read and correct rather than a crash.

The tools

Security, and why it's two layers deep

query_database is the tool most likely to be handed something destructive, so it has two independent guards:

1. SQLite's own read-only mode. The connection is opened as file:store.db?mode=ro, so a write is refused by the database engine — not by my code.

2. Statement validation. Only SELECT and WITH…SELECT are accepted. Multiple statements, PRAGMA, ATTACH, and every DDL/DML keyword are rejected with a message explaining why. SQL comments are stripped first, so a forbidden keyword can't hide behind --.

search_files and read_file resolve every path and confirm it is still inside the configured root. ../../../../etc/passwd, absolute paths, and symlinks pointing outside are all refused.

All of this is covered by tests — see below.

Setup

git clone https://github.com/harshhh817/mcp-server-demo.git
cd mcp-server-demo
pip install -r requirements.txt
python make_sample_db.py     # builds data/store.db

Verify it works before wiring it into anything:

python test_server.py

Claude Code

From inside the repo:

claude mcp add toolkit -- python /absolute/path/to/mcp-server-demo/server.py

Or commit a .mcp.json at your project root so the whole team gets it:

{
  "mcpServers": {
    "toolkit": {
      "command": "python",
      "args": ["/absolute/path/to/mcp-server-demo/server.py"],
      "env": {
        "TOOLKIT_DB_PATH": "/absolute/path/to/mcp-server-demo/data/store.db",
        "TOOLKIT_FILES_ROOT": "/absolute/path/to/mcp-server-demo/sandbox"
      }
    }
  }
}

Then check it registered:

claude mcp list

Claude Desktop

Edit claude_desktop_config.json:

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

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

{
  "mcpServers": {
    "toolkit": {
      "command": "python",
      "args": ["/absolute/path/to/mcp-server-demo/server.py"],
      "env": {
        "TOOLKIT_DB_PATH": "/absolute/path/to/mcp-server-demo/data/store.db",
        "TOOLKIT_FILES_ROOT": "/absolute/path/to/mcp-server-demo/sandbox"
      }
    }
  }
}

Restart Claude Desktop. The tools appear under the tools icon in the chat input.

Two things that trip people up: paths must be absolute — MCP servers don't inherit your shell's working directory. And "command": "python" must be a Python that has mcp installed; if you use a virtualenv, point at /path/to/venv/bin/python instead.

Configuration

Point these at your own database and folder and the server is immediately useful — nothing in the tool code is specific to the sample data.

Try it

Once connected, ask Claude things like:

What tables are in the database?

Which customer has spent the most, and on what?

Find every file mentioning connection pooling and summarise the issue.

What's the weather in Delhi this week?

Claude picks the tool, writes the SQL, and reads the result. For the second question it typically calls describe_database, then issues a three-table join on its own.

Tests

test_server.py spawns the server as a subprocess and drives it as a real MCP client over stdio — the same transport Claude Desktop uses. It isn't testing the functions in-process; it's testing the server.

python test_server.py

Real output:

connected to 'toolkit' v1.0.0

tool discovery
  ok    server advertises 5 tools  (describe_database, get_weather, query_database, read_file, search_files)
  ...
query_database - rejected queries
  ok    DROP is rejected  (only SELECT (or WITH ... SELECT) queries are allowed)
  ok    stacked statements rejected  (only one statement per call; remove the extra ';')
  ok    comment-hidden DROP rejected  (only one statement per call; remove the extra ';')
  ok    database still intact after attacks
path traversal is refused
  ok    ../ escape refused  ('../server.py' resolves outside the allowed root)
  ok    absolute path refused  ('/etc/passwd' resolves outside the allowed root)
get_weather (live network call)
  ok    resolves the location  (Delhi, National Capital Territory of Delhi, India)
  ok    returns a temperature  (29.2degC, overcast)

40 passed, 0 failed

Tech

Python 3.10+ · mcp v2 (official SDK) · SQLite · stdlib urllib (no HTTP dependency for the weather tool)

server.py            tool definitions + MCPServer wiring
toolkit_mcp/
├── database.py      read-only SQL, validation, schema introspection
├── files.py         sandboxed search and read
└── weather.py       Open-Meteo client
make_sample_db.py    seeded sample database
test_server.py       end-to-end test over real stdio MCP transport
sandbox/             what the file tools are allowed to see

Note on SDK versions: mcp v2 renamed FastMCP to MCPServer (from mcp.server.mcpserver import MCPServer). Most tutorials online still show the v1 FastMCP import, which raises ModuleNotFoundError on v2. This repo targets v2.

Building one for your stack

The pattern generalises. A tool is a decorated function with a clear description and typed arguments — the description is what Claude reads to decide whether to call it, so it's worth writing carefully. Common asks I can build:

• Query your internal Postgres/MySQL read replica safely

• Search and summarise a documentation or ticket archive

• Call your company's REST API with auth handled server-side

• Read and write to a Google Sheet or Notion database

Hire me

I build MCP servers and Claude Code workflows — tested, documented, and delivered fast.

• Fiverr: My Fiverr profile

• Upwork: My Upwork profile

• GitHub: github.com/harshhh817

Final-year B.Tech CSE · AWS Certified Cloud Practitioner · Delhi, India

MIT licensed — see LICENSE.