MCP 101

Calling a tool

0. Make sure that nothing is listening on ports 8000 and 8080. Open 3 generously sized terminals on your screen.
1. Download a sensible model. Qwen 3.5 4B is sensible.
2. Compile fresh llama.cpp:

   git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
   cmake -B build && cmake --build build --config Release -j 6
   

3. Launch the llama in terminal #1:

   ./llama-server -m ~/Downloads/Qwen3.5-4B-Q8_0.gguf --ctx-size 4096 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.00 --verbose --webui-mcp-proxy
   

4. Clone this repository:

   https://github.com/behavioral-ds/mcp-example && cd mcp-example
   

5. Install deps: poetry install && poetry shell
6. Launch MCP in terminal #2: python mcp_serve.py
7. Execute the Agentic Call™ in terminal #3: python call.py
8. Observe the dance between LLM <-> Inference engine <-> MCP <-> Client.

Using MCP prompts

9. Open llama web UI at http://localhost:8080/, go to settings and add a new MCP server: <img width="585" src="https://github.com/user-attachments/assets/b4fedffa-550f-4f1c-af80-79e2c5876826" /> <img width="585" src="https://github.com/user-attachments/assets/9fc83783-9725-4e68-9396-a158d6bf9335" />

10. Select "MCP prompt" when drafting a new message:<br /> <img width="241" src="https://github.com/user-attachments/assets/12495540-ffcc-4710-a345-c941fd05bafd" />

11. That's your @mcp.prompt() parsed into UI element, click it:<br /> <img width="257" src="https://github.com/user-attachments/assets/29d82c3d-d133-47b8-ad42-e0a2b00ae5af" />

12. ...and supply some meaningful content:<br /> <img width="220" src="https://github.com/user-attachments/assets/e4a9b415-31ba-4386-a6cb-24c509229c29" />

13. Then click "Use prompt" and rejoice: <img width="803" alt="image" src="https://github.com/user-attachments/assets/7c7ccaf8-d62a-4b6f-93a0-fa62ae4a0a8a" />