fastmcp-sample-server 0.4.0

A demo Model Context Protocol (MCP) server built on FastMCP — 15 tools, 5 resources, 3 prompts, SQLite-backed notes, per-call telemetry, and a Streamlit UI.

Sample Local MCP Server

A minimal Model Context Protocol server built with FastMCP. It exposes a handful of demo tools (math, random number, server time, text analysis, echo) over SSE on http://localhost:8000, so you can wire it into an MCP-aware client such as VS Code Copilot Chat, Claude Desktop, or any custom agent.

Project structure

sample-mcp-server-script/
├── server/                  # The MCP server
│   ├── __init__.py
│   ├── __main__.py          # enables `python -m server`
│   ├── app.py               # FastMCP app + tools + resources + prompts
│   └── notes_db.py          # SQLite-backed notes store
├── clients/                 # Ways to talk to the server
│   ├── cli.py               # CLI: list / call / smoke
│   └── streamlit_app.py     # Streamlit UI
├── tests/                   # pytest suite
│   ├── test_notes_db.py
│   └── test_server.py
├── docs/                    # Misc notes
│   ├── instructions.txt
│   └── testing_server.txt
├── data/                    # Local SQLite DB + logs (gitignored)
├── Dockerfile
├── .dockerignore
├── .gitignore
├── README.md
└── requirements.txt

Tools exposed

Tool	Description
calculate	Evaluate a safe math expression, e.g. 2 + 2 * 10.
get_server_time	Return the current server time (UTC).
random_number	Random int between min_val and max_val.
analyze_text	Word / character / sentence stats for a text blob.
echo	Echoes a message back — handy for connectivity testing.
hash_text	MD5/SHA1/SHA256/SHA512 hash of a string.
base64_encode	Base64-encode a string (standard or URL-safe alphabet).
base64_decode	Decode a base64 string back to UTF-8 text.
uuid_generate	Generate one or more UUIDs (v1 or v4).
password_generate	Cryptographically strong random password with configurable charset.
json_format	Pretty-print and validate a JSON string.
regex_match	Find all matches of a regex pattern in text (first 50, with groups).
convert_units	Convert length / weight / temperature between common units.
fetch_url	HTTP GET/HEAD an http(s) URL and return status, headers, truncated body.
weather	Current weather for a lat/lon via the free Open-Meteo API (no key).
note_add	Create a new note (title + body) in the local SQLite store.
note_list	List recent notes, newest first. Optional substring search.
note_get	Fetch a single note by id.
note_update	Update a note's title and/or body.
note_delete	Delete a note by id.

Resources exposed

URI	Description
resource://server/info	Server metadata: name, version, uptime, notes count.
resource://notes/all	All notes (max 500), newest first.
resource://notes/{note_id}	A single note by id, e.g. resource://notes/3.
resource://telemetry/recent	Last 100 tool calls (tool, ts, duration_ms, status).
resource://telemetry/summary	Aggregates: 60-min KPIs + 24-h per-tool counts.

Prompts exposed

Prompt	Description
summarize_text	Summarize a text blob in a chosen style.
code_review	Structured code review (bugs / security / improvements).
explain_error	Plain-language explanation of an error + likely fixes.

Prerequisites

• Python 3.10+ (3.11 recommended)
• pip and venv available

Setup

From the project root:

# 1. Create and activate a virtual environment
python3 -m venv .venv
source .venv/bin/activate

# 2. Install dependencies
pip install --upgrade pip
pip install -r requirements.txt

Run the server

# With the venv activated:
python -m server

You should see FastMCP start an SSE server on port 8000. The SSE endpoint is:

http://localhost:8000/sse

Leave this terminal running while you use the server from a client.

---

Quick-start cheat sheet

Three terminals, all from the project root after source .venv/bin/activate.

Terminal 1 — start the server

python -m server
# or with debug logging to a file
python -m server --log-level DEBUG --log-file data/mcp.log

Terminal 2 — Streamlit UI

streamlit run clients/streamlit_app.py
# opens http://localhost:8501

Terminal 3 — CLI client

python -m clients.cli list                    # discover tools
python -m clients.cli smoke                   # run the built-in smoke suite
python -m clients.cli call echo --args '{"message":"hi"}'

---

CLI flags

python -m server --help

# Examples
python -m server --transport stdio
python -m server --transport sse --host 127.0.0.1 --port 9000
python -m server --log-level DEBUG --log-file data/mcp.log

All flags can also be set via environment variables: MCP_TRANSPORT, MCP_HOST, MCP_PORT, MCP_LOG_LEVEL, MCP_LOG_FILE, MCP_NOTES_DB.

Verify it's running

In another terminal:

curl -N http://localhost:8000/sse

You should get a streaming response (press Ctrl+C to stop). If the connection opens and stays open, the server is healthy.

Test from the terminal

Two convenient clients are included.

1. CLI harness — clients/cli.py

source .venv/bin/activate

# List every tool the server exposes
python -m clients.cli list

# Call any tool (arguments are a JSON object)
python -m clients.cli call echo --args '{"message": "hello"}'
python -m clients.cli call password_generate --args '{"length": 32, "use_symbols": true}'
python -m clients.cli call weather --args '{"latitude": 28.6139, "longitude": 77.2090}'

# Run a small predefined smoke suite that exercises ~9 tools
python -m clients.cli smoke

# Point at a non-default URL
python -m clients.cli --url http://localhost:9000/sse list

2. Streamlit UI — clients/streamlit_app.py

A zero-config web UI that lists tools, auto-renders a form from each tool's input schema, and shows results + call history.

source .venv/bin/activate
streamlit run clients/streamlit_app.py

Streamlit opens at http://localhost:8501. In the sidebar, click Connect / Refresh to load tools from http://localhost:8000/sse (editable), pick a tool from the dropdown, fill in the form, hit Call tool.

3. Raw curl flow (advanced)

If you want to test the JSON-RPC wire protocol directly, see docs/instructions.txt for the manual SSE + /messages/?session_id=… flow.

Use it from VS Code Copilot Chat

A ready-to-use config is included at .vscode/mcp.json:

{
  "servers": {
    "local-sample-mcp": {
      "type": "sse",
      "url": "http://localhost:8000/sse"
    }
  }
}

Steps:

1. Make sure the server is running (python -m server).
2. Open this folder in VS Code.
3. Open Copilot Chat and switch to Agent mode.
4. Click the tools picker — the local-sample-mcp server and its tools should appear.
5. Ask something like: "Use the calculate tool to compute 12 * (3 + 4)" or "Call analyze_text on this paragraph."

Use it from Claude Desktop

Claude Desktop currently launches MCP servers via stdio, so the easiest path is to use mcp-proxy to bridge stdio → SSE. Add an entry like this to your Claude Desktop config:

{
  "mcpServers": {
    "local-sample-mcp": {
      "command": "mcp-proxy",
      "args": ["http://localhost:8000/sse"]
    }
  }
}

Then start the Python server in a terminal as shown above.

Switching to stdio transport (optional)

If you prefer running over stdio (no port needed):

python -m server --transport stdio

And update your client config to launch the script directly, e.g.:

{
  "servers": {
    "local-sample-mcp": {
      "type": "stdio",
      "command": "python",
      "args": ["-m", "server"],
      "cwd": "/absolute/path/to/sample-mcp-server-script"
    }
  }
}

Telemetry

Every tool call is recorded to a local SQLite file (data/telemetry.db by default) via the @traced decorator in server/telemetry.py. Per call we store: timestamp, tool name, duration in ms, and status (ok or error) — and the exception message when something raises. We deliberately do not record arguments or return values to keep secrets, tokens, hashed inputs, password outputs, etc. out of the telemetry DB.

Tool-level failures returned as {"error": "..."} count as ok (they are a normal Python return); only uncaught exceptions are recorded as error.

Toggle via env vars:

Variable	Default	Meaning
MCP_TELEMETRY	1	Set to 0 to disable recording.
MCP_TELEMETRY_DB	data/telemetry.db	Path to the telemetry SQLite file.

The Streamlit UI has a 📊 Telemetry tab that pulls resource://telemetry/summary and resource://telemetry/recent and renders: KPI cards (calls, error rate, avg + p50 / p95 / p99 latency), a Plotly bar chart of calls per tool (last 24 h), slowest-calls and recent-errors tables side-by-side, a latency scatter over time, and a filterable recent-calls table. An auto-refresh (5 s) checkbox keeps it live, and the Tool caller tab shows a per-tool "called N times, avg X ms" badge under the description.

Guardrails

The server applies a small set of defensive checks so a careless (or malicious) agent can't trivially hang the process, escape the network sandbox, or fill up the database. They live in server/safety.py and are wired into the relevant tools.

Tool	Guardrail	Why
calculate	AST-based evaluator instead of eval(). Only + - * / // % **, parens, unary +/-, numeric literals. Exponent capped at 100; expression capped at 200 chars.	eval() is dangerous even with a charset filter — 9**9**9**9 would peg a CPU core forever.
fetch_url	Resolves the hostname and rejects loopback, private (10/8, 172.16/12, 192.168/16), link-local (incl. 169.254/16 cloud metadata), multicast, and reserved IPs. Streams the response and stops after max_bytes.	Blocks SSRF and prevents downloading huge files just to truncate them.
regex_match	Pattern capped at 1 KB, text at 1 MB, execution wrapped in a 2-second timeout via a worker thread.	Stops catastrophic backtracking (ReDoS) from blocking the server.
note_add / note_update	Title required, ≤200 chars; body ≤100 KB; both validated at the DB layer.	Prevents one bad call from bloating the SQLite file.

What's not included (yet, by design — this is a local demo):

• No per-tool rate limiting
• No authentication on the SSE/HTTP transports
• No output-size cap on tool return values
• No prompt-injection sanitization on fetched HTML

If you ever expose this beyond localhost, add at least auth and rate limits before doing so.

Run with Docker

# Build the image
docker build -t sample-mcp-server .

# Run with a persistent notes DB volume
docker run --rm -p 8000:8000 -v "$PWD/data:/data" sample-mcp-server

# Override transport/host/port/log-level via env or CLI args
docker run --rm -p 9000:9000 -e MCP_PORT=9000 sample-mcp-server
docker run --rm sample-mcp-server --transport stdio

The image exposes port 8000 and stores the notes DB at /data/notes.db (mount a volume there to persist it across container restarts).

Run the tests

source .venv/bin/activate
pytest -q

tests/test_notes_db.py covers the SQLite store and tests/test_server.py covers every pure tool, resource, and prompt by importing the server module directly.

Troubleshooting

• ModuleNotFoundError: No module named 'fastmcp' — activate the venv and rerun pip install -r requirements.txt.

• Port 8000 already in use — change the port=8000 argument in server/app.py, and update the url in .vscode/mcp.json to match.

• VS Code doesn't see the server — restart the MCP server from the Copilot Chat tools picker, or reload the VS Code window.

• calculate returns an error — only 0-9, +, -, *, /, (, ), ., and spaces are accepted (intentional safety restriction).

  ---

  Demo Video

  https://github.com/user-attachments/assets/3dacf24b-3717-4b82-a036-e5ff4131bd09