MCP servers are just REST APIs in a polite wrapper - here's 5 lines of Python

If you've been watching the MCP (Model Context Protocol) ecosystem from the sidelines, here's a quietly important detail: a lot of MCP servers are also just plain REST APIs underneath. The MCP layer is a polite wrapper that says "Claude, here are tools you can call." But the underlying HTTP endpoints are right there, ready to be called from requests.get(...) like any other JSON API.

That matters because the most interesting MCP servers are useful even if you've never opened Claude Desktop or Cursor. You can drop them into a Streamlit app, a Jupyter notebook, a Lambda function, a Discord bot, an Airflow DAG, or a cron job. The MCP integration is gravy on top.

I'll show this with a concrete example: pulling per-symbol ML option fair values and 31-dimension news-bias scores into pandas in 5 lines.

The setup

I run Helium MCP, which started as an MCP server and recently grew a plain REST surface. Both speak the same data:

• Per-symbol ML options pricing - predicted fair value, probability ITM, Greeks
• 31-dimension news-bias scoring across 3.2M articles and 5,000 sources
• Real-time market data, top trading strategies, semantic meme search, source bias profiles

The MCP endpoint is https://heliumtrades.com/mcp. The REST endpoints live under https://heliumtrades.com/ with paths like /mcp_search/, /mcp_option_price/, /mcp_ticker/, /mcp_url_bias/. 50 free queries per IP. No signup, no API key needed for the free tier.

Five lines of Python

import requests

r = requests.get(
    "https://heliumtrades.com/mcp_search/",
    params={"q": "apple earnings", "limit": 3},
    timeout=30,
)
print(r.json())

You get back a JSON list of articles with full bias scoring across all 31 dimensions per article: credibility, sensationalism, overconfidence, opinion_vs_fact, scapegoating, ai_authorship_probability, covering_responses, oversimplification, and 23 more.

Loading into pandas

This is where it gets fun. The JSON is already flat enough that pandas just works:

import pandas as pd, requests

resp = requests.get(
    "https://heliumtrades.com/mcp_search/",
    params={"q": "federal reserve", "limit": 50},
)
df = pd.json_normalize(resp.json())
print(df[["source", "credibility", "sensationalism", "opinion_vs_fact"]].head())

Now you can do everything pandas does: groupby source, compute mean credibility, plot a credibility-vs-sensationalism scatter, filter to high-AI-authorship-probability articles, etc.

Option fair values, also 5 lines

import requests

r = requests.get(
    "https://heliumtrades.com/mcp_option_price/",
    params={
        "symbol": "AAPL",
        "strike": 200,
        "expiration": "2026-06-19",
        "option_type": "call",
    },
)
print(r.json())
# {'predicted_price': 20.64, 'prob_itm': 0.52, 'delta': 0.55, 'gamma': 0.02, 'vega': 0.41, ...}

You get back a model-derived fair value and prob_ITM next to market price. The diff between the two is a (testable, scorable) prediction.

A small Streamlit app

Once the API returns JSON, building a Streamlit app is essentially a wrapper exercise:

import streamlit as st, requests, pandas as pd

q = st.text_input("Search query", "tariffs")
limit = st.slider("Results", 1, 50, 10)

if st.button("Go"):
    resp = requests.get(
        "https://heliumtrades.com/mcp_search/",
        params={"q": q, "limit": limit},
    )
    df = pd.json_normalize(resp.json())
    st.dataframe(df[["title", "source", "credibility", "sensationalism", "ai_authorship_probability"]])
    st.bar_chart(df.groupby("source")["credibility"].mean())

This is the smallest realistic media-bias dashboard I've ever written. It's about 12 lines.

Where this fits in your stack

The point isn't that this one API is special. The point is that MCP servers with REST surfaces are a quietly powerful new class of API. They are:

• LLM-native by design - built so an LLM can call them without a custom integration
• Schema-rich - the MCP tool spec doubles as auto-generated API documentation
• Free-tiered aggressively - because the operator wants discoverability in LLM clients
• Composable from anything HTTP - Python, JS, curl, Go, n8n, Zapier, Make

If you're a data scientist who's never installed Claude Desktop and never wants to: that's fine. Treat MCP servers as a directory of unusually well-curated free REST APIs and start with the ones that solve a problem you already have.

For finance and news intelligence specifically, the full Helium MCP REST endpoint list is:

The MCP server config (for Cursor / Claude Desktop / Windsurf) is:

{ "mcpServers": { "helium": { "url": "https://heliumtrades.com/mcp" } } }

But honestly - if Python is your thing - just open a notebook and requests.get. The whole point of a public REST surface is that you don't have to care about anything else.

Source, schema, full tool spec: github.com/connerlambden/helium-mcp. Page: heliumtrades.com/mcp-page.

If you build something with it, I'd love to see it. Open an issue, or send a notebook - happy to feature good demos.