MCP Server: Documentation Retrieval & Web Scraping (uv + FastMCP)

This project provides a minimal, async MCP (Model Context Protocol) server that exposes a tool for retrieving and cleaning official documentation content for popular AI / Python ecosystem libraries. It uses:

• fastmcp to define and run the MCP server over stdio.
• httpx for async HTTP calls.
• serper.dev for Google-like search (via API).
• groq API (LLM) to clean raw HTML into readable text chunks.
• python-dotenv for environment variable management.
• uv as the package manager & runner (fast, lockfile-based, Python 3.11+).

Features

• Search restricted to official docs domains (uv, langchain, openai, llama-index).
• Tool: get_docs(query, library) returns concatenated cleaned sections with SOURCE: labels.
• Streaming-safe async design (chunking large HTML pages before LLM cleaning).
• Separate client.py demonstrating how to connect as an MCP client and call the tool, then post-process with an LLM.

---

Quick Start

Prerequisites:

• Python 3.11+
• uv installed (https://docs.astral.sh/uv/)
• API keys for: SERPER_API_KEY, GROQ_API_KEY

1. Clone & Install

git clone <your-repo-url> mcp-server-python
cd mcp-server-python
uv sync

This will create/refresh a .venv based on pyproject.toml + uv.lock.

2. Environment Variables

Create a .env file in the project root:

SERPER_API_KEY=your_serper_api_key_here
GROQ_API_KEY=your_groq_api_key_here

Optional: add other model settings if you later extend functionality.

3. Run the MCP Server Directly

uv run mcp_server.py

The server will start and wait on stdio (no extra output unless you add logging). It registers the tool get_docs.

4. Use the Provided Client

uv run client.py

You should see something like:

Available tools: ['get_docs']
ANSWER: <model-produced answer referencing SOURCE lines>

If the list is empty, ensure the server started correctly and no exceptions were raised (add logging—see below).

---

Tool: get_docs

Signature:

get_docs(query: str, library: str) -> str

Supported libraries (keys): uv, langchain, openai, llama-index.

Flow:

1. Build a site-restricted query: site:<docs-domain> <query>.
2. Call Serper API for organic results.
3. Fetch each result URL (async) via httpx.
4. Split HTML into ~4000‑char chunks (memory safety & LLM limits).
5. Clean each chunk using Groq LLM (openai/gpt-oss-20b) with a system prompt.
6. Concatenate and label each block with SOURCE: <url> for traceability.

Returned value: A large text blob suitable for retrieval-augmented prompting, preserving source attribution lines.

---

Architecture

File overview:

File	Purpose
mcp_server.py	Defines FastMCP instance and implements search_web, fetch_url, and the get_docs tool.
client.py	Launches server via stdio, lists tools, calls get_docs, then feeds result to an LLM for a user-friendly answer.
utils.py	HTML cleaning helper (currently uses LLM + trafilatura for extraction and Groq for chunk transformation).
.env	Environment variables (excluded from VCS).
pyproject.toml	Declares dependencies and metadata.
uv.lock	Reproducible lockfile generated by uv.

---

Dependency Notes

Core runtime deps (from pyproject.toml):

• fastmcp – MCP server helper.
• httpx – async HTTP client.
• groq – Groq API client.
• python-dotenv – load variables from .env.
• trafilatura – heuristic content extraction (currently partially used / can be extended).

---

Logging & Debugging

To see what the server is doing, you can temporarily add:

import logging, sys
logging.basicConfig(level=logging.INFO, stream=sys.stderr)

Place near the top of mcp_server.py after imports. Since protocol uses stdout for JSON-RPC, send logs to stderr only.

Common issues:

• Empty tool list: The server exited early or crashed—add logging.
• SERPER_API_KEY missing → 401 or empty search results.
• GROQ_API_KEY missing → LLM cleaning fails (exception in get_response_from_llm).
• Network timeouts: Adjust timeout in httpx.AsyncClient calls.

---

Extending

Ideas:

• Add caching layer (e.g., sqlite or in-memory dict) to avoid re-fetching same URLs.
• Parallelize URL fetch + clean with asyncio.gather() (mind rate limits / LLM cost).
• Add another tool (e.g., summarize_diff, list_endpoints).
• Provide structured JSON output (list of sources + cleaned text) instead of concatenated string.
• Add tests using pytest + pytest-asyncio (mock Serper + LLM APIs).

---

Example Programmatic Use (Without Client Wrapper)

If you want to call the tool directly in a Python script using the client-side MCP library:

from mcp.client.stdio import stdio_client
from mcp import ClientSession, StdioServerParameters
import asyncio

async def demo():
	params = StdioServerParameters(command="uv", args=["run", "mcp_server.py"])
	async with stdio_client(params) as (r, w):
		async with ClientSession(r, w) as session:
			await session.initialize()
			tools = await session.list_tools()
			print([t.name for t in tools.tools])
			docs = await session.call_tool("get_docs", {"query": "install", "library": "uv"})
			print(docs.content[:500])

asyncio.run(demo())

---

Running With Active Virtualenv

If you have an already activated virtual environment and want to use that instead of the project’s pinned environment, you can force uv to target it:

uv run --active client.py

Otherwise, uv will warn that your active $VIRTUAL_ENV differs from the project .venv but continue using the project environment.

---

Troubleshooting Cheat Sheet

Symptom	Cause	Fix
No tools listed	Server not running / crashed	Add stderr logging; run uv run mcp_server.py manually
AttributeError on .text	Cleaner returned None	Ensure you return actual string from fetch_url / LLM call
401 from Serper	Bad/missing API key	Check .env and reload shell
Empty search results	Narrow query	Simplify query or verify domain key
High latency	Many sequential LLM chunk calls	Batch or reduce chunk size

---

Acknowledgments

• https://www.youtube.com/@AI.with.Hassan
• Serper.dev for search API
• Groq for fast OSS model serving
• Astral for uv
• MCP ecosystem for protocol foundation