abhi-nexus

Building Multi-Agent Hybrid System with MCP Sever

Abhi-Nexus MCP Server — Project Documentation

Author: Abhi
Built: June 2026
Stack: Python, FastMCP, Docker, Unraid, Ollama, VS Code

What Is Abhi-Nexus?

Abhi-Nexus is a production-grade autonomous news intelligence system. It fetches articles from 20+ tech sources every day, ranks them using AI, generates summaries in English and Hindi, and delivers a curated email digest at 8:15 AM — fully automatically, with zero human intervention.

It is engineered as a portfolio project that demonstrates real-world multi-agent architecture, production systems thinking, and cost-conscious data engineering.

Related MCP server: NewsDigest MCP

What Is MCP?

MCP (Model Context Protocol) is an open standard that gives LLMs the ability to interact with external systems in real time. Think of it as giving an AI hands — it can fetch data, trigger actions, and query systems beyond its training data.

Without MCP, an LLM can only think and talk. With MCP, it can act.

Real-world analogy: MCP is like a USB standard. Before MCP, every AI integration was a custom cable that only worked with one specific system. After MCP, it becomes a USB port — any compatible client can plug in and use it.

What We Built

We exposed Abhi-Nexus as an MCP server — making the pipeline accessible to any MCP-compatible client (VS Code, Cursor, Open WebUI, or custom agents) without changing any existing agent code.

Architecture

MCP Client (VS Code + Continue + Local LLM)
                    |
         MCP Protocol (SSE, port 7070)
                    |
         Abhi-Nexus MCP Server
         (mcp_server.py — FastMCP 2.0.0)
                    |
    ┌───────────────┼───────────────┐
    ▼               ▼               ▼
Fetch Agent    Rank Agent     Delivery Agent
(RSS feeds)   (GPT scoring)   (Email SMTP)
    ▼               ▼               ▼
raw_articles   ranked_articles   Email at
   .json           .json          8:15 AM

Deployment Stack

MCP Tools Exposed

MCP Resources

File Structure Added

z:/abhi-nexus/
└── mcp_server.py          ← New: MCP server exposing all tools

Changes to Existing Files

Bugs Fixed During Build

Three pre-existing bugs were discovered and fixed during the MCP integration:

Bug 1 — OpenAI proxies Error

Symptom: Client.__init__() got an unexpected keyword argument 'proxies'
Cause: openai==1.30.0 was too old, incompatible with newer httpx
Fix: Upgraded to openai==1.56.0

Bug 2 — JSON Truncation in Rank Agent

Symptom: Unterminated string parse errors, all articles falling back to score 0
Cause: max_tokens=1500 was too low — GPT response cut off mid-JSON
Fix: Increased to max_tokens=6000, shortened summary prompt to 1 sentence

Bug 3 — Unicode Garbage Characters

Symptom: Stack Overflow articles had invisible Unicode characters (\u200b, \ufeff etc.) inflating token count
Cause: Stack Overflow RSS feed embeds tracking/watermark characters in content
Fix: Added character filter in fetch_agent.py clean_text() function

How to Connect

VS Code (Continue Extension)

Add to ~/.continue/config.yaml:

name: Abhi-Nexus Config
version: 1.0.0
schema: v1

models:
  - name: DeepSeek 14B
    provider: ollama
    model: deepseek-r1:14b
    apiBase: http://localhost:11434

mcpServers:
  - name: abhi-nexus
    url: http://192.168.4.46:7070/sse

Cursor IDE

Add to Cursor MCP settings:

{
  "mcpServers": {
    "abhi-nexus": {
      "url": "http://192.168.4.46:7070/sse"
    }
  }
}

Verify Server is Running

curl http://192.168.4.46:7070/sse

Expected response:

event: endpoint
data: /messages/?session_id=...
: ping - 2026-06-13 ...

Adding New Tools

Adding a new MCP tool is one decorated function:

@mcp.tool()
def scan_vulnerabilities(target: str = "dependencies") -> dict:
    """Scan Abhi-Nexus for security vulnerabilities."""
    # your logic here
    return {"vulnerabilities": [...]}

FastMCP automatically generates the schema, validates inputs, and registers the tool. No other changes needed.

Interview Talking Points

What you built:

"I exposed a multi-agent autonomous news pipeline as an MCP server running on self-hosted Docker on Unraid. Any MCP-compatible client — VS Code, Cursor, or a custom agent — can now trigger the pipeline, query ranked articles, and monitor health in real time. The LLM stays local using Ollama, so there is zero cloud dependency for the AI layer."

Why MCP over function calling:

"Function calling is per-model and per-request — you define tools inline for one specific app. MCP is transport-agnostic and model-agnostic. I build the server once and any MCP client can use it forever without rewriting the integration."

Architecture decision — why not rewrite agents as MCP servers:

"The agents already work. MCP is an additive interface layer, not a replacement. I wrapped the existing FastAPI surface and Python functions as MCP tools without touching agent logic. This is the same pattern Meta uses — expose capabilities through a standard interface without disrupting the underlying system."

Cost discipline:

"The pipeline runs at $2/month. Batching 10 articles per API call instead of 1 saves 75% on GPT costs. The MCP server itself is free — open protocol, self-hosted, no per-query charges."

What MCP Is Not

Next Steps

• Add security vulnerability scanning tool (scan_dependencies, check_cve)

• Upgrade to FastMCP 3.x for Streamable HTTP transport (replaces SSE)

• Add authentication layer to MCP server (API key header)

• Build a second agent that calls Abhi-Nexus as a tool in a larger workflow

• Add deep research tool — top 3 articles → web search → enriched summary