Raj Aryan

Posted on Jun 14

10 Killer MCP Projects to Supercharge Your AI Engineering Portfolio

#mcp #startup #programming #javascript

🚀 Learning AI isn't just theory — it's about building real tools. Fast.

In this post, I’m giving you 10 hands-on MCP (Model Context Protocol) project ideas you can run entirely locally. Forget cloud lock-in — these are fast, private, and seriously cool.

Each project includes:

🔧 What it does
💡 Why it’s worth building
🧪 Code you can run immediately

Perfect for engineers, indie hackers, and tinkerers looking to level up their local-first AI dev game.

🧠 Wait, What’s MCP?

MCP is like a personal assistant API bridge — it lets local AI apps like Cursor or Claude Desktop interact with your local tools, databases, files, and scripts.

You say: “Check this stock trend” → MCP talks to your local CSV + code → Returns chart/insight → No cloud, no lag, full control.

1. ✉️ Lightweight MCP Socket Client

The foundation.
Build a socket-based local MCP client that can send & receive queries.

import socket, json

def mcp():
    s = socket.socket()
    s.connect(('localhost', 9999))
    s.send(json.dumps({"action": "status"}).encode())
    print(json.loads(s.recv(1024).decode()))
    s.close()

mcp()

Next: Add a Flask or FastAPI backend.

2. 🔁 Dual-Source RAG Engine

Search smarter.
Local document retrieval that uses web scraping as a fallback.

from llama_index import VectorStoreIndex, SimpleDirectoryReader
import requests
from bs4 import BeautifulSoup

docs = SimpleDirectoryReader("docs").load_data()
index = VectorStoreIndex.from_documents(docs)

def query(q):
    answer = index.as_query_engine().query(q)
    return fallback(q) if answer.score < 0.4 else str(answer)

def fallback(q):
    r = requests.get(f"https://lite.search/?q={q}")
    return BeautifulSoup(r.text, 'html.parser').find('p').text

print(query("What is AGI?"))

3. 📊 Local Finance Insight Bot

Analyze local stock CSVs with moving averages and charts.

import pandas as pd
import matplotlib.pyplot as plt

def analyze(ticker):
    df = pd.read_csv(f"data/{ticker}.csv")
    df['MA30'] = df['Close'].rolling(30).mean()
    plt.plot(df['Date'], df['Close'], label='Close')
    plt.plot(df['Date'], df['MA30'], label='MA30')
    plt.legend(); plt.show()
    return f"{ticker} is {'rising' if df['Close'].iloc[-1] > df['MA30'].iloc[-1] else 'falling'}"

print(analyze("TSLA"))

4. 🎙️ Speak-to-Search Assistant

Voice interface + RAG = your own Alexa

import speech_recognition as sr
from llama_index import SimpleDirectoryReader, VectorStoreIndex

r = sr.Recognizer()
index = VectorStoreIndex.from_documents(SimpleDirectoryReader("notes").load_data())

def listen():
    with sr.Microphone() as source:
        audio = r.listen(source)
    return r.recognize_google(audio)

def respond():
    q = listen()
    return str(index.as_query_engine().query(q))

print(respond())

5. 🧠 Unified Data Query Server (MindsDB)

Ask anything about your local databases.

from flask import Flask, request, jsonify
import mindsdb_sdk

app = Flask(__name__)
mdb = mindsdb_sdk.connect()

@app.route('/query', methods=['POST'])
def query():
    q = request.json['q']
    return jsonify(mdb.query(q).fetch().to_dict())

app.run(port=5000)

6. 📓 Memory Bridge for Claude + Cursor

Store key-value memory across tools using SQLite.

import sqlite3

def setup():
    with sqlite3.connect('mem.db') as conn:
        conn.execute('CREATE TABLE IF NOT EXISTS store (k TEXT PRIMARY KEY, v TEXT)')

setup()

def set_note(k, v):
    with sqlite3.connect('mem.db') as con:
        con.execute('REPLACE INTO store VALUES (?, ?)', (k, v))

def get_note(k):
    with sqlite3.connect('mem.db') as con:
        r = con.execute('SELECT v FROM store WHERE k=?', (k,)).fetchone()
    return r[0] if r else 'none'

set_note('focus', 'Work on MCP today')
print(get_note('focus'))

7. 📄 RAG for Messy PDFs

Extract text from messy PDFs using PyMuPDF.

import fitz
from llama_index import Document, VectorStoreIndex

def extract(file):
    text = ''.join([page.get_text() for page in fitz.open(file)])
    return VectorStoreIndex.from_documents([Document(text=text)]).as_query_engine()

print(extract("sample.pdf").query("What's discussed?"))

8. 🧪 Synthetic Table Generator

Generate test data locally using SDV.

from sdv.single_table import GaussianCopulaSynthesizer
from flask import Flask, request, jsonify
import pandas as pd

app = Flask(__name__)

@app.route('/generate', methods=['POST'])
def generate():
    df = pd.read_csv(request.json['file'])
    synth = GaussianCopulaSynthesizer(df)
    synth.fit()
    return jsonify(synth.sample(100).to_dict())

app.run(port=5050)

9. 🔍 Local Research Agent Crew

Multi-agent CrewAI for deep local research.

from crewai import Agent, Task, Crew
from llama_index import VectorStoreIndex, SimpleDirectoryReader

docs = VectorStoreIndex.from_documents(SimpleDirectoryReader("src").load_data())
researcher = Agent("Analyst", "Break down info", tools=[docs.as_query_engine()])
task = Task("Summarize AI + ethics", agent=researcher)
crew = Crew([researcher], [task])
print(crew.kickoff())

10. 🔐 Local Auth Without Passwords

Build your own authentication system — no APIs.

from flask import Flask, request, jsonify

app = Flask(__name__)
users = {"raj": "faceprint123"}

@app.route('/login', methods=['POST'])
def login():
    body = request.json
    return jsonify({"success": users.get(body['user']) == body['auth']})

app.run(port=3001)

🚀 Final Thoughts

These MCP projects are powerful learning tools and excellent portfolio pieces. You’ll gain:

🧠 Local AI deployment skills
🔧 Real-world use case experience
🔒 Cloud-free, privacy-first AI setups

💬 Let me know which project you want to build — or improve together!

#ai #machinelearning #mcp #openai #cursorai #crewai #llamaindex #privacyfirst #buildinpublic

DEV Community