mcp-drone-agent

Give Claude actual flight control over a drone via the Model Context Protocol. This is the demo project for the "drones" video: an LLM agent reads a plain-English mission, plans a search pattern itself, and calls real MCP tools (takeoff, move_to, scan, land, ...) to fly it.

Why a simulator instead of a real flight stack

The paper this is inspired by ("Taking Flight with Dialogue," PX4-based Drone Agent, arXiv 2506.07509) uses PX4 + ROS 2 + NVIDIA Isaac Sim, which needs a GPU Linux box and a fair amount of setup. This project keeps the same idea -- natural language in, tool calls out, flight path as the payoff -- but swaps the flight stack for a small deterministic Python simulator (drone_sim.py) so it runs anywhere in about 10 seconds with no GPU, no ROS2, no simulator install. The MCP server (mcp_server.py) is real, not a mock -- swap drone_sim.World for a PX4/MAVSDK backend later and the tool layer above it doesn't change.

Related MCP server: ArduPilot MCP Server Sandbox

How it's wired together

mission text  -->  agent.py  --(MCP stdio)-->  mcp_server.py  -->  drone_sim.py
                      |                                                |
                 LLM (Ollama / Claude)                          physics / world
                 decides which tool                             state / flight log
                 to call next

• drone_sim.py -- a 60x60m field, three static obstacles, three hidden targets (one is the mission objective, two are decoys), simple kinematics, battery drain, and a flight event log. Pure Python, no dependencies.

• mcp_server.py -- wraps the simulator in five MCP tools using the official MCP Python SDK (FastMCP), served over stdio.

• agent.py -- the MCP client. --mode ollama (default) uses a local LLM via Ollama. --mode mock is a scripted lawnmower search (no LLM). --mode claude uses Anthropic's API.

• visualize.py -- renders the flight log into a top-down PNG: field boundary, obstacles, the flown path, detections, obstacle detours, home/landing markers.

• demo_mission.py -- runs the whole pipeline in one command.

Setup

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

Local LLM (recommended)

Install Ollama, pull a tool-capable model, and keep ollama serve running:

ollama pull qwen2.5:7b

Default model is qwen2.5:7b (strong tool calling). Also works with llama3.1:8b / llama3.2 if already pulled. Prefer models whose Ollama card lists tools support.

Cloud Claude (optional)

For --mode claude, copy .env.example to .env and set ANTHROPIC_API_KEY. agent.py loads .env via python-dotenv.

Live web console (recommended)

Same teal theme / sidebar layout as medical-assistant. Backend streams telemetry over WebSocket; the map animates the drone in real time.

# Terminal 1 — API
cd backend
pip install -r requirements.txt
uvicorn app.main:app --reload --port 8000

# Terminal 2 — UI
cd frontend
npm install
npm run dev

Open http://localhost:3000/mission, pick Mock or Ollama, hit Start mission.

CLI demo

# Local LLM via Ollama (default) -- real agent, no Anthropic bill
python demo_mission.py --mode ollama

# Scripted search, no LLM -- proves MCP + sim only
python demo_mission.py --mode mock

# Cloud Claude (needs credits / API key)
python demo_mission.py --mode claude

# Pick another local model
python demo_mission.py --mode ollama --model llama3.1:8b

All modes write flight_logs/latest.jsonl (raw event log), flight_logs/agent_transcript.json (every tool call + LLM reasoning text, if any).

Same seed (--seed 42 by default) = same field layout every run, so you can re-record a take without the obstacle/target positions moving around on you.

Point Claude Desktop at it directly

Since mcp_server.py is a real MCP server, you can also add it straight to Claude Desktop's config and fly missions from the chat window instead of the script:

{
  "mcpServers": {
    "drone-agent": {
      "command": "python3",
      "args": ["/absolute/path/to/mcp-drone-agent/mcp_server.py"]
    }
  }
}

Restart Claude Desktop, then just type a mission in chat, e.g. "take off, search the field, and land next to the survivor marker."