Multi-Agent Research Assistant & MCP Server

This project implements the assignment in two small parts:

1. A LangGraph multi-agent workflow with one Supervisor and two specialist workers.

2. A FastMCP server with two tools and a client that calls those tools.

The language model is Ollama, so no OpenAI API key is required.

1. What The Project Demonstrates

The Supervisor receives a question and chooses the appropriate worker:

User question
		 |
		 v
Supervisor Agent
		 |------------------------------|
		 v                              v
Research Agent                 Analysis Agent
		 |                              |
		 v                              v
Knowledge-base tool             Comparison tool

For a question that needs both workers, the Supervisor asks the Research Agent for evidence first, then gives that evidence to the Analysis Agent.

The MCP part is independent of the LangGraph part:

MCP Client ---> FastMCP Server
										|------ get_weather(city)
										|------ get_news(topic)

Related MCP server: AIE8-MCP Server

2. Prerequisites

• Windows PowerShell

• Python 3.10 or newer

• Ollama

• An Ollama model such as llama3.2

The repository already contains the requested virtual environment at multivenv.

3. Installation

Open PowerShell in the project directory:

cd D:\LLMEngg_8-Multi-Agent-Research-Assistant-MCP-Server
.\multivenv\Scripts\Activate.ps1
python -m pip install -r requirements.txt
Copy-Item .env.example .env

If PowerShell does not allow activation, use the virtual environment directly for every command:

.\multivenv\Scripts\python.exe -m pip install -r requirements.txt

Do not use a different system Python. Otherwise imports such as langchain_core may be missing.

4. Configure Ollama

Start Ollama in a separate terminal. If Ollama is already running as a desktop application, skip ollama serve.

ollama serve
ollama pull llama3.2

The default configuration in .env is:

OLLAMA_MODEL=llama3.2
OLLAMA_BASE_URL=http://localhost:11434

You can use another model, but it must support chat and tool calling well enough for LangGraph. For example:

OLLAMA_MODEL=qwen2.5:7b

5. Run The Multi-Agent Assistant

Run a research-only question:

python main.py "What does MCP do?"

This should route to the Research Agent, which searches data/knowledge_base.txt.

Run an analysis question:

python main.py "Compare these two ideas: solar power uses sunlight; wind power uses moving air."

This should route to the Analysis Agent, which uses compare_texts.

Run a collaboration question:

python main.py "Research solar and wind power, then compare their main trade-offs."

This question needs both workers. The expected workflow is:

1. The Supervisor calls ask_research_agent.

2. The Research Agent calls search_knowledge_base.

3. The Supervisor sends the evidence to ask_analysis_agent.

4. The Analysis Agent calls compare_texts.

5. The Supervisor returns one final answer.

The default question can also be run without an argument:

python main.py

6. Run The MCP Demonstration

The client calls both FastMCP tools:

python -m mcp_client.client

Expected output is similar to:

Weather: Cloudy, 15 C
News: Mock headline: New developments in artificial intelligence are being monitored by the research team.

The client uses FastMCP's in-process client transport so the demonstration is reliable and easy to run locally. It still uses the real MCP client/server protocol. The standalone server can be started for an MCP-compatible host with:

python mcp_server/server.py

That standalone server uses MCP stdio transport.

7. Run Tests

python -m pytest -q

The tests cover the deterministic tools without requiring Ollama or a model download. The LangGraph construction can also be checked without making a model request:

$env:OLLAMA_MODEL = "llama3.2"
python -c "from agents.supervisor import build_supervisor; build_supervisor(); print('Supervisor created')"

8. File Structure

agents/
	model.py             Shared ChatOllama configuration
	research_agent.py    Research worker created with create_react_agent
	analysis_agent.py    Analysis worker created with create_react_agent
	supervisor.py        Supervisor and wrapped worker tools

tools/
	research_tool.py     Local knowledge-base lookup tool
	analysis_tool.py     Structured two-text comparison tool

data/
	knowledge_base.txt   Local evidence used by the Research Agent

mcp_server/
	server.py            FastMCP server and its two tools

mcp_client/
	client.py            Client demonstration calling both MCP tools

main.py                Command-line entry point for the Supervisor
tests/                  Deterministic tool tests
requirements.txt        Python dependencies
.env.example            Ollama configuration template

9. Assignment Objective Checklist

10. Troubleshooting

No module named langchain_core

The system Python is being used. Activate multivenv or use the direct interpreter path:

.\multivenv\Scripts\python.exe main.py "What does MCP do?"

connection refused from Ollama

Start Ollama and confirm the model exists:

ollama serve
ollama list
ollama pull llama3.2

The model does not call tools

Use a tool-capable chat model, keep the question explicit, and try the collaboration example from Section 5. Small or older models may answer directly without using a tool.

The knowledge base returns no answer

The Research Agent only searches the local file. Add more paragraphs to data/knowledge_base.txt and rerun the question.

11. Important Scope Note

This is a simple educational implementation. The knowledge base is a mock local data source, weather and news are mock MCP results, and the LLM routing is tested manually with Ollama. The deterministic tools are covered by automated tests.