🔌 AI MCP Toolkit

An MCP (Model Context Protocol) server that exposes AI capabilities — two of which call other running services in this portfolio over HTTP, and one self-contained — as standardized tools any MCP-compatible client can use directly, including Claude Desktop.

🎯 What It Does

MCP standardizes how an AI client calls external code. Write a tool once as an MCP server, and any MCP-compatible client — Claude Desktop, Claude Code, Cursor — can discover and call it, with no client-specific integration work.

This server exposes three tools, two of which are thin clients calling other repos' running FastAPI services over HTTP — not reimplementations of similar logic:

web_research   → HTTP call to ai-research-agent's /research/ endpoint (port 8003)
review_pr       → HTTP call to ai-pr-reviewer's /pr-review/ endpoint (port 8004)
explain_concept  → self-contained, calls Groq directly — genuinely new, no reuse claim

The distinction matters and is worth being precise about: web_research and review_pr have no logic of their own. If ai-research-agent or ai-pr-reviewer isn't running, those tools fail outright — they don't fall back to anything. That failure mode is the proof that this is real cross-service interconnection, not two repos that happen to do similar things.

📸 Screenshots

Running server.py

MCP Inspector — tool schemas and live testing MCP Inspector Local dev tool showing all 3 tools auto-discovered from @mcp.tool() decorators, with their generated input/output schemas.

Connected and running in Claude Desktop Claude Desktop connection Settings → Developer → Local MCP servers, showing ai-toolkit with status running.

A real tool call inside a Claude conversation Tool call in action Claude recognizing a request matches the web_research tool, invoking it, and returning a result grounded in live search — not its own training data.

Running inter_server_communication.py

Without MCP Inspector — PR REVIEW MCP Inspector Claude made a normal call to view the public PR and based on diff gives the suggestion.

With MCP Inspector - PR REVIEW MCP Inspector Now claude made custom mcp tool calls and forwarded the request to ai-pr-reviewer over HTTP and returned the identical structured result.

Side by side, these two screenshots are the actual evidence of cross-service reuse — same backend, same output, two different ways of reaching it.

✨ Features

• Protocol-standard tool exposure — built on the official MCP Python SDK (FastMCP)
• Real cross-repo interconnection — web_research and review_pr are HTTP clients of other services in this portfolio, isolated in their own module for clarity
• Auto-generated schemas — tool input/output schemas derive from Python type hints and docstrings
• Honest dependency, not duplication — if a backing service is down, its tool fails; nothing is silently reimplemented as a fallback
• Client-agnostic — works with Claude Desktop, Claude Code, MCP Inspector, or any future MCP client unchanged

🏗️ Architecture

Claude Desktop (or any MCP client)
        │  stdio + MCP protocol
        ▼
server.py  (FastMCP — tool registration, schema generation)
        │
        ├── explain_concept ──────────────► Groq directly (no other service)
        │
        └── inter_server_communication.py
                ├── web_research ─── HTTP ──► ai-research-agent  (port 8003)
                └── review_pr     ─── HTTP ──► ai-pr-reviewer     (port 8004)

inter_server_communication.py is a separate module specifically because it carries the cross-service dependency — keeping it isolated from server.py makes the "this tool depends on another repo being up" relationship explicit and easy to point to, rather than buried inside tool definitions.

🧠 How It Works

server.py registers tools with @mcp.tool(). Two of those tools don't contain business logic — they import functions from inter_server_communication.py, which makes an HTTP POST to another repo's running FastAPI service and returns its response, reshaped into a readable string for the MCP client.

# inter_server_communication.py
def call_research_agent(topic: str, depth: str = "quick") -> str:
    response = httpx.post(
        f"{RESEARCH_AGENT_URL}/research/",
        json={"topic": topic, "depth": depth},
        timeout=120
    )
    response.raise_for_status()
    data = response.json()
    findings = "\n".join(f"- {f}" for f in data.get("key_findings", []))
    return f"{data.get('summary', '')}\n\nKey findings:\n{findings}"

# server.py
@mcp.tool()
def web_research(topic: str, depth: str = "quick") -> str:
    """Run the ai-research-agent service's autonomous web research agent on a topic."""
    return call_research_agent(topic, depth)

When Claude calls review_pr with a GitHub PR URL, the request goes: Claude → MCP server → inter_server_communication.py → HTTP → ai-pr-reviewer's /pr-review/ endpoint → GitHub API (to fetch the diff) → Groq (to generate the review) → back through the same chain to Claude. Five hops, three repos, one conversational request.

🗂️ Project Structure

ai-mcp-toolkit/
├── server.py                        # Tool registration, FastMCP entry point
├── inter_server_communication.py    # HTTP clients for ai-research-agent and ai-pr-reviewer
├── .env.example
└── .gitignore

🚀 Getting Started

Prerequisites

• Python 3.11+
• Groq API key — free
• Tavily API key — free
• ai-research-agent and ai-pr-reviewer repos, runnable locally

Installation

git clone https://github.com/vyavahare-kishor/ai-mcp-toolkit
cd ai-mcp-toolkit

curl -LsSf https://astral.sh/uv/install.sh | sh
uv venv
source .venv/bin/activate
uv add "mcp[cli]" httpx groq python-dotenv

Configuration

cp .env.example .env

GROQ_API_KEY=your_groq_api_key_here
RESEARCH_AGENT_URL=http://localhost:8003
PR_REVIEWER_URL=http://localhost:8004

Run the dependent services first

# Terminal 1 — ai-research-agent
cd ai-research-agent && uvicorn main:app --reload --port 8003

# Terminal 2 — ai-pr-reviewer
cd ai-pr-reviewer && uvicorn main:app --reload --port 8004

Test locally — MCP Inspector

uv run mcp dev server.py

# for cross server testing
uv run mcp dev inter_server_communication.py

Try review_pr with a real public GitHub PR URL while watching ai-pr-reviewer's terminal — you should see the incoming request logged there, confirming the call actually crossed into that repo.

Connect to Claude Desktop

{
  "mcpServers": {
    "ai-toolkit": {
      "command": "uv",
      "args": ["--directory", "/absolute/path/to/ai-mcp-toolkit", "run", "server.py"]
    }
  }
}

# for cross server testing
{
  "mcpServers": {
    "ai-toolkit": {
      "command": "uv",
      "args": ["--directory", "/absolute/path/to/ai-mcp-toolkit", "run", "inter_server_communication.py"]
    }
  }
}

Restart Claude Desktop, then ask: "Use review_pr to review this PR: [github PR url]"

🗺️ Roadmap

• [ ] Wrap ai-customer-support-bot's /support/ask as a 4th cross-service tool
• [ ] Add a fallback message (not silent failure) when a dependent service is unreachable
• [ ] Add an MCP resource exposing recent research/review history
• [ ] Authentication for remote deployment beyond local stdio

🔗 Related Projects

Part of an AI-native engineering portfolio. Full journey: ai-engineering-journey

👨‍💻 Author

Kishor Vyavahare Senior Software Engineer → AI Native Engineer

11+ years of backend engineering (Ruby on Rails, PostgreSQL, AWS). Now building production AI systems — RAG pipelines, agents, multi-agent crews, and protocol-standard tool exposure with real cross-service architecture.

📄 License

MIT License — use it, fork it, build on it.